IJACSA Volume 16 Issue 6 - thesai.org

IJACSA Volume 16 Issue 6

Copyright Statement: This is an open access publication licensed under a Creative Commons Attribution 4.0 International License, which permits unrestricted use, distribution, and reproduction in any medium, even commercially as long as the original work is properly cited.

View Full Issue

Paper 1: Detecting and Preventing Money Laundering Using Deep Learning and Graph Analysis

Abstract: Money laundering is a major worldwide issue facing financial organizations, with its increasingly complicated and changing methods. Conventional rule-based anti-money laundering (AML) systems can fail to identify advanced fraudulent activity. This study shows a new hybrid model to detect suspicious transaction patterns precisely by efficiently combining GraphSAGE, a graph-based Machine Learning (ML) technique, with Long Short-Term Memory (LSTM) networks. The suggested approach uses GraphSAGE's relational capabilities for graph-structured anomaly detection and the temporal strengths of LSTM for sequence modeling. With excessive traditional ML and stand-alone Deep Learning (DL) techniques, the Hybrid LSTM-GraphSAGE model achieves an accuracy of 95.4% using a simulated dataset reflecting real-world financial transactions. The findings show how well our combined strategy lowers false positives and improves the identification of advanced AML operations. This work opens the path for creating real-time, intelligent, flexible money laundering detection systems appropriate for current financial situations.

Author 1: Mamunur R Raja

Author 2: Md Anwar Hosen

Author 3: Md Farhad Kabir

Author 4: Sharmin Sultana

Author 5: Shah Ahammadullah Ashraf

Author 6: Rakibul Islam

Keywords: Anti-money laundering (AML); deep learning (DL); LSTM; GraphSAGE; graph analysis; transaction monitoring; hybrid fusion model

Paper 2: Transforming the Working Style of Call Center Agents Through Generative AI

Abstract: As Generative Artificial Intelligence (Gen AI) is evolving rapidly, there is a significant change in the approach by the contact center industry with respect to work culture. Historically, customer service agents working in a contact center used to depend significantly on static scripts and fragmented information systems, which thereby resulted in delayed resolutions, cognitive overload, and made them deliver inconsistent customer experiences. This study explores the paradigm shift that's occurring in contact service centers through implementing Gen AI. Real-time intent recognition, contextual response generation, and personalized engagement across channels are some novel capabilities introduced by Gen AI by adopting Large Language Models (LLMs). Organizations can reduce Average Handling Time (AHT), improve First Contact Resolution (FCR), and enhance Customer Satisfaction (CSAT) scores by integrating Gen AI into core workflows such as issue summarization, behavioral analytics, sentiment tracking, and knowledge retrieval. In order to demonstrate the quantifiable improvements in agent performance and customer engagement, this study adopted a blended research design by combining enterprise case studies, simulation scenarios, and comparative KPI evaluations. Furthermore, it addresses implementation bottlenecks such as onboarding efficiency, multilingual support, emotional intelligence, and real-time guidance. With reference to the industry standards, ethical considerations such as data privacy, algorithmic bias, and explainability are examined. Case examples that are collected from the industry leaders are leveraged to validate the study's conclusions. Through this study, a structured and well-organized roadmap for enterprises is delivered, which aims at transforming contact centers from reactive service units into proactive, intelligence-driven ecosystems.

Author 1: Satya Karteek Gudipati

Keywords: Generative AI; call center transformation; agent augmentation; LLMs; sentiment analysis; hyper-personalization; conversational AI; AI ethics

Paper 3: AI-Driven Education: Integrating Machine Learning and NLP to Transform Child Learning Systems

Abstract: An Artificial Intelligence-driven child learning system with a Machine Learning and Natural Language Processing-based approach to dynamically personalize educational experiences for children is proposed in this study. Using a Sentence-BERT model to encode student queries for the computation of semantic similarity and knowledge domains to be retrieved. A T5-based transformer model writes verbose, personalized feedback, and a Gradient Boosting Machine classifier predicts the appropriate learning outcomes. The content difficulty and personalization of educational trajectories across content are set by an integrated adaptive learning engine that monitors and adjusts for student performance. On the General Knowledge QA dataset, classification accuracy reaches 85.2%, and the ROC-AUC score is 0.912, which has been proven to be reliable in real-world cases. It also produces positive effects regarding the understanding and preference for learners of adaptive systems, as observed in user studies. AI technologies have exciting potential to deliver scalable, personalized education for young learners, as demonstrated in this work.

Author 1: Masuma Akter Semi

Author 2: Md Borhan Uddin

Author 3: Sharmin Sultana

Author 4: Motmainna Tamanna

Author 5: Azim Uddin

Author 6: Khandakar Rabbi Ahmed

Keywords: Artificial intelligence; machine learning; natural language processing; adaptive learning systems; sentencebert; gradient boosting machine; personalized feedback

Paper 4: Application of Blockchain Frameworks for Decentralized Identity and Access Management of IoT Devices

Abstract: The growth in IoT devices means an ongoing risk of data vulnerability. The transition from centralized ecosystems to decentralized ecosystems is of paramount importance due to security, privacy, and data use concerns. Since the majority of IoT devices will be used by consumers in peer-to-peer applications, a centralized approach raises many issues of trust related to privacy, control, and censorship. Identity and access management lies at the heart of any user-facing system. Blockchain technologies can be leveraged to augment user authority, transparency, and decentralization. This study proposes a decentralized identity management framework for IoT environments using Hyperledger Fabric and Decentralized Identifiers (DIDs). The system was simulated using Node-RED to model IoT data streams, and key functionalities including device onboarding, authentication, and secure asset querying were successfully implemented. Results demonstrated improved data integrity, transparency, and user control, with reduced reliance on centralized authorities. These findings validate the practicality of blockchain-based identity management in enhancing the security and trustworthiness of IoT infrastructures.

Author 1: Sushil Khairnar

Keywords: Blockchain; decentralization; identity and access management; ethereum; hyperledger

Paper 5: Location Based Augmented Reality Navigation Application

Abstract: This paper presents a novel Augmented Reality (AR) navigation system to overcome limitations of conventional 2D map-based applications in advanced real-world environments. Current AR navigation systems solutions often lack dynamic adaptation to user behavior and fail to deliver context-aware, personalized guidance. Addressing these gaps, we present a markerless, location-based AR system integrating three innovations: 1) a Dynamic Predictive Navigation module with Long Short-Term Memory (LSTM) networks for anticipating user intention and dynamically optimizing routes in real time; 2) a Smart POI Ranking system with sentiment analysis, live user feedback, and social media trends for presenting personalized and context-aware recommendations; and 3) a 3D AR interface built with Unity and ARCore for enhancing spatial understanding and reducing cognitive burden through visually engaging guidance. Experimental evaluation presents improved navigation responsiveness, reduced rerouting effort, and increased user interaction with recommended POIs. This work contributes a scalable and adaptive solution towards real-time AR navigation, with applicability to smart city mobility and context-aware spatial computing.

Author 1: Samridhi Sanjay Pramanik

Author 2: Aishwary Pramanik

Keywords: Augmented reality; location based application; markerless AR; GPS based application; real world augmentation; unity; ARCore; navigation system; user interaction

Paper 6: Metabolite Screening for Heart Disease Using Support Vector Machine-Based AI

Abstract: Algorithms for feature selection are growing in interest among researchers aiming to connect specific features in a dataset with specific classifications. Recent developments in machine learning, particularly Support Vector Machine-based artificial intelligence algorithms have demonstrated excellent classification performance in highly nonlinear data. However, identifying which features contribute most to classification re-mains challenging, especially when datasets include hundreds of variables. Initially, features must be screened to narrow down the set for deeper analysis. Metabolomics datasets are one such case, where many features must be examined to determine those associated with heart disease diagnosis. This work applies a Genetic Algorithm, incorporating a penalized likelihood approach with Support Vector Machines for mutation, to stochastically search the feature space. A large-scale simulation study demonstrates that the proposed method achieves a high true feature identification rate while maintaining a reasonable false identification rate. The method is then applied to a Qatar BioBank dataset focused on heart disease, reducing the number of candidate metabolites from 232 to 37.

Author 1: Edward L. Boone

Author 2: Ryad A. Ghanam

Author 3: Faten S. Alamri

Author 4: Elizabeth B. Amona

Keywords: Machine learning; genetic algorithm; support vector machines; classification; heart disease; metabolites

Paper 7: Enhancing Deepfake Content Detection Through Blockchain Technology

Abstract: Deepfake technology poses a growing threat to the authenticity and trustworthiness of digital media, necessitating the development of advanced detection mechanisms. While AI-based methods have shown promise, they generally face limitations in terms of generalization and scalability. We present a blockchain-enabled watermarking technique, characterized by its immutable, transparent, and decentralized nature, which offers a robust complementary approach for enhancing media authentication through methods such as cryptographic watermarking, decentralized identity, and content provenance tracking. To train and evaluate blockchain-based watermarking and deepfake detection systems, a variety of large-scale datasets are utilized. Video datasets include UADFV (49 real, 49 fake), Deepfake-TIMIT (320 real, 640 fake), DFFD (1000 real, 3000 fake), Celeb-DF v2 (590 real, 5639 fake), DFDC (23,564 real, 104,500 fake), DeeperForensics-1.0 (50,000 real, 10,000 fake), FaceForensics++ (1000 real, 5000 fake), and ForgeryNet (99,630 real, 121,617 fake). Image datasets include DFFD (58,703 real, 240,336 fake), FFHQ (70,000 GAN-generated), iFakeFaceDB (87,000 fake), 100k AI Faces, and over 2.8 million samples in ForgeryNet. Despite integration challenges such as scalability, computational cost, and standardization, blockchain-based solutions show promise in tracking content origin and enhancing verification. Simulation results demonstrate that the proposed blockchain-enabled watermarking achieves a higher accuracy in detecting fake content compared to existing machine learning methods.

Author 1: Qurat-ul-Ain Mastoi

Author 2: Muhammad Faisal Memon

Author 3: Salman Jan

Author 4: Atif Jamil

Author 5: Muhammad Faique

Author 6: Zeeshan Ali

Author 7: Abdullah Lakhan

Author 8: Toqeer Ali Syed

Keywords: Blockchain; deep fake; convolutional neural network (CNN); long short-term memory (LSTM); RNN (recurrent neural network); video and image

Paper 8: Advancements in Deep Learning for Malaria Detection: A Comprehensive Overview

Abstract: Malaria remains a critical global health issue, with millions of cases reported annually, particularly in resource-limited regions. Timely and accurate diagnosis is vital to ensure effective treatment, reduce complications, and control transmission. Conventional diagnostic methods, including microscopy and Rapid Diagnostic Tests (RDTs), face considerable limitations such as dependency on skilled personnel, limited sensitivity at low parasitemia levels, and cost constraints. In response, deep learning technologies—especially Convolutional Neural Networks (CNNs)—have emerged as promising tools to overcome these barriers by enabling automated diagnostics based on medical imaging, significantly enhancing precision and scalability. This paper presents a comprehensive review of recent advances in deep learning for malaria diagnosis, highlighting the role of publicly available datasets in driving innovation. It analyzes leading architectures—such as ResNet, VGG, and YOLO—based on their classification performance, including accuracy, sensitivity, and computational efficiency. Furthermore, the review discusses novel directions such as mobile-integrated diagnostics and multi-modal data fusion, which can enhance diagnostic accessibility in low-resource settings. Despite notable progress, challenges remain in terms of dataset imbalance, lack of generalizability, and barriers to clinical deployment. The paper concludes by outlining future research directions and emphasizing the need for robust, adaptable models that can support global malaria control and eradication strategies.

Author 1: Kiswendsida Kisito Kabore

Author 2: Desire Guel

Author 3: Flavien Herve Somda

Keywords: Malaria detection; deep learning; Convolutional Neural Networks (CNNs); medical imaging; automated diagnostics

Paper 9: Application of Deep Learning-Based Image Compression Restoration Technology in Power System Unstructured Data Management

Abstract: In power-system unstructured-data management, a large volume of images from inspection drones, substation cameras, and smart meters is heavily compressed due to bandwidth and storage constraints, resulting in lower resolution that hinders defect detection and maintenance decisions. Although deep-learning super-resolution (SR) techniques have made significant advances, real-world deployments still require a balance between reconstruction accuracy and model lightweightness. To meet this need, we introduce a channel-attention-embedded Transformer SR method (CAET). The approach adaptively injects channel attention into both the Transformer’s global features and the convolutional local features, harnessing their complementary strengths while dynamically enhancing critical information. Tested on five public datasets and compared with six representative algorithms, CAET achieves the best or second-best performance across all upscaling factors; at 4× enlargement, it outperforms the advanced SwinIR method by 0.09 dB in PSNR on Urban100 and by 0.30 dB on Manga109, with noticeably improved visual quality. Experiments demonstrate that CAET delivers high-precision, low-latency restoration of compressed images for the power sector while keeping model complexity low.

Author 1: Junjie Zha

Author 2: Aiguo Teng

Author 3: Xinwen Shan

Author 4: Hao Tang

Author 5: Zihan Liu

Keywords: Image compression; attention mechanism; multimodal fusion; unstructured data in the power industry; image data

Paper 10: Impact of Auxiliary Information in Generative Artificial Intelligence Models for Cross-Domain Recommender Systems

Abstract: Recommender systems (RSs) are significant in enhancing the experiences of users across different online platforms. One of the major problems faced by the conventional RSs is difficulties in getting precise preferences for users, mostly for the users that has limited previous interaction data, and this eventually affects the performance of the conventional techniques to solve the data sparsity problem. To address this challenge, this study proposes an Auxiliary-Aware Conditional GAN (AUXIGAN) model that integrates heterogeneous auxiliary information into both the generator and discriminator networks to enhance representation learning to enhance the performance of the cross-domain recommender systems (CDRS). Most researchers consider only the rating matrix of users-items and ignore the impact of auxiliary information on the interaction functions, which is very significant to the recommendation accuracy to solve data sparsity problems. The proposed novel technique considers features concatenation, attention-based fusion networks, contrastive representation learning, knowledge transfer, and multi-modal embedding alignment techniques to enhance the user-item interaction matrix. Our experiments on benchmark datasets show that the proposed model significantly outperformed state-of-the-art RSs models, the key metrics utilized are: RMSE, Precision, Recall, and MAE, which show the influence of incorporating auxiliary information into the GAN-based CDRS. In conclusion, the integration of auxiliary information on generative adversarial networks models represents a substantial advancement in the field of CDRS, and the results of the proposed models on two real-world datasets show that the proposed model significantly outperforms collaborative filtering and other GAN-based techniques.

Author 1: Matthew O. Ayemowa

Author 2: Roliana Ibrahim

Author 3: Noor Hidayah Zakaria

Author 4: Yunusa Adamu Bena

Keywords: Generative adversarial networks; auxiliary information; cross-domain recommender systems; data sparsity; knowledge transfer

Paper 11: The Anomaly Detection Algorithm Based on Random Matrix Theory and Machine Learning

Abstract: This study focuses on anomaly detection algorithms. Aiming at the limitations of traditional methods in complex data processing, an innovative algorithm that integrates random matrix theory and machine learning is proposed. First, different types of data, such as numerical values, texts, and images, are preprocessed, and random matrices are constructed. Hidden abnormal features are mined through specific transformations and then classified by optimized machine learning models. In the experimental stage, multiple data sets, such as KDD Cup 99, are selected to compare with classic algorithms such as DBSCAN and Isolation Forest. The results show that the innovative algorithm has a detection accuracy of 95%, a recall rate of 93%, and an F1 value of 94% on the KDD Cup 99 data set, which is significantly improved compared with the comparison algorithm. It also performs well on other data sets, with an average accuracy increase of seven percentage points and a recall rate increase of eight percentage points. The results demonstrate that the proposed algorithm can effectively mine data anomaly patterns, achieve efficient and accurate anomaly detection in complex data sets, and provide strong support for applications in related fields.

Author 1: Yongming Lu

Keywords: Random matrix theory; machine learning; anomaly detection; experimental simulation

Paper 12: SkinDiseaseXAI: XAI-Driven Neural Networks for Skin Disease Detection

Abstract: Accurate classification of skin diseases is an important step toward early diagnosis and therapy. However, deep learning models are frequently used in therapeutic contexts without transparency, reducing confidence and acceptance. This study introduces SkinDiseaseXAI, a unique convolutional neural network (CNN) that uses Grad-CAM++ to classify ten different types of skin diseases and provide visual explanations. The proposed model was trained using a publicly available dataset of dermatoscopic images following preprocessing and augmentation. SkinDiseaseXAI achieved 76.12% training accuracy and 66.25% validation accuracy in 20 epochs. We used Grad-CAM++ to generate heatmaps that highlighted discriminative regions inside the lesion areas, thereby improving interpretability. The experimental results indicate that the model has the ability not only to perform multi-class skin disease categorization but also to provide interpretable visual outputs, which improves the transparency and dependability of decision-making processes. This concept has the possibility to improve clinical diagnosis by merging performance and explainability.

Author 1: Ammar Nasser Alqarni

Author 2: Abdullah Sheikh

Keywords: XAI; skin disease; Grad-CAM++; convolutional neural networks; clinical interpretability; melanoma; eczema; atopic dermatitis; fungal infections

Paper 13: Emotion Recognition Algorithm Based on Multi-Modal Physiological Signal Feature Fusion Using Artificial Intelligence and Deep Learning

Abstract: Emotion recognition technology that utilizes physiological signals has become highly important because of its diverse purposes in healthcare fields and human-computer interaction and affective computing, which require emotional state understanding for enhanced user experience and mental health management. Support Vector Machines (SVM) and Random Forest (RF) serve as traditional machine learning approaches for emotion classification, but they struggle to accurately model spatial, temporal and long-range dependencies within multimodal physiological data, which leads to degraded overall performance. Created an Attention-Based CNN-BiLSTM-Transformer Model, which unites several neural network structures to extract features and classify information more effectively. This model implements Convolutional Neural Networks for detecting spatial patterns at the raw level of numerous physiological signals, which contain Electroencephalography, Electrocardiography, Galvanic Skin Response, and Electromyography. BiLSTM works as a temporal model which analyzes time-series physiological patterns through dual-directional contextual processing to create improved features from historical data patterns. The Transformer encoder serves to detect extended relationships between sequence items for better emotional change comprehension throughout time. The classification accuracy receives additional improvement because an attention-based fusion mechanism applies dynamic importance weights to different physiological signals, so the most significant features optimize the ultimate decision process. Testing of the proposed model using publicly accessible DEAP and AMIGOS resulted in 88.2% accuracy on DEAP while achieving 89.5% accuracy on AMIGOS, and both outcomes exceeded conventional machine learning methods as well as baseline deep learning approaches, which used CNN-LSTM and Transformer-only models. Testing showed that the attention mechanism successfully determined how to weigh multiple features, which resulted in better classification success. A deep learning framework based on TensorFlow and PyTorch operates throughout the implementation in Python to provide an efficient solution for emotion recognition in physiological signals.

Author 1: Yue Pan

Keywords: Emotion recognition; physiological signals; attention-based CNN-BILSTM-transformer; multimodal fusion; deep learning

Paper 14: Dataset Development for Classifying Kick Types for Martial Arts Athletes Using IMU Devices

Abstract: This study aims to establish a dataset of kicks in Kempo martial arts to categorize athletes' kick types based on their movement patterns. The real problem addressed in this research is the lack of an accurate, efficient, and portable system to automatically recognize and classify kick types in martial arts training, especially outside of controlled laboratory environments. Previous studies often relied on optical motion capture systems, which, while accurate, are expensive and impractical for real-world training settings. To overcome this limitation, this study utilizes wearable motion sensing technology and analyzes the data with an Inertial Measurement Unit (IMU) sensor. During data collection, IMU sensors are attached to athletes to monitor their movements during training or competitions. In this research, an IMU sensor type MPU6050 is used, controlled by an ESP32 microcontroller, and the data is collected via wireless communication to a data collection server. The dataset comprises gyroscope readings for angular velocity and accelerometer measurements for linear acceleration in three axes. The study evaluates three kick types: straight kicks, sidekicks, and roundhouse kicks. It employs machine learning methodologies utilizing three principal classification algorithms: Support Vector Machine (SVM), k-Nearest Neighbors (k-NN), and Random Forest (RF). These algorithms were selected due to their distinct strengths in processing sensor data derived from the accelerometer and gyroscope embedded within the IMU device. The findings indicate that the SVM algorithm successfully identified and categorized kick types in Shorinji Kempo martial arts athletes with a 96.7 per cent accuracy rate when using a dataset of 70 samples and two sensors. However, when three sensors were used, the accuracy decreased to approximately 92.4 per cent. In contrast, the k-NN algorithm achieved a classification accuracy of 92.4 per cent with a dataset of 70 samples, k = 3, and three sensors. Analyzing the contributions of features to classification provides in-depth insight into the key characteristics of movement patterns for kick type recognition.

Author 1: Rudy Gunawan

Author 2: Suhardi

Author 3: Widyawardana Adiprawita

Author 4: Tommy Apriantono

Keywords: Classification; dataset; machine learning; inertial sensors; IMU; martial art; sport; motion capture

Paper 15: Comparative Analysis of Deep Learning Techniques for Passive Underwater Acoustic Target Recognition: Overview, Challenges, and Future Directions

Abstract: Passive underwater acoustic target recognition (UATR) involves analyzing acoustic waves captured by passive sonar to extract valuable information about submerged targets. The underwater acoustics community has increasingly turned its attention to deep learning techniques, owing to their remarkable success in image recognition tasks. This study presents a comprehensive overview of the evolution of UATR techniques, categorizing them into three distinct groups: early methods, conventional machine learning approaches, and modern deep learning-based techniques. Additionally, it provides an in-depth summary of the recognition process utilizing deep learning, detailing various deep network architectures, classifiers specifically designed for underwater acoustic target recognition, and different data input modalities. Finally, the study synthesizes current research findings and outlines potential future directions for advancements in this field, emphasizing opportunities for innovation across these three categories.

Author 1: Song Yifei

Author 2: Mohamad Farhan Mohamad Mohsin

Keywords: Underwater acoustic target recognition; deep learning; deep network architecture; classifier

Paper 16: Method for Tea Leaf Plucking Timing Prediction with High Resolution of Images Based on YOLO11

Abstract: As a method for estimating the time when tea leaves reach their peak quality (amino acid content) (optimum picking time), our previous study revealed that the optimum picking time is when the accumulated temperature from the detection of germination of new buds reaches 600°C. However, the accuracy of this germination detection was insufficient, so the estimation accuracy of the optimum picking time was also insufficient. Since annotation accuracy is extremely important for germination detection by YOLO11, strict attention is paid to annotation by hand and by increasing the number of training datasets. The detection accuracy has been improved compared to the germination detection by YOLOv8, which was previously proposed and used relatively low-resolution images. The conclusion of this study is that the estimation method of the optimum picking time based on the criterion that the optimum picking time (amino acid content reaches its peak) is effective when the accumulated temperature from germination detection meets the condition of 600°C. The effectiveness of this method has been confirmed by comparison with germination detection by experts. For tea farmers, being able to predict the optimum picking time, when the amino acid content in the new buds is at its peak, is important, and we are sure it will have a positive impact on agricultural researchers studying this subject.

Author 1: Kohei Arai

Author 2: Yoho Kawaguchi

Keywords: Tealeaf plucking; YOLO; budding detection; spatial resolution; optical image; annotation; germination rate

Paper 17: The Factors Influencing Internet of Things Adoption in Public Hospitals: A Pilot Study

Abstract: Incorporating Internet of Things (IoT) technologies in healthcare represents a significant leap forward, capable of transforming the delivery and management of medical services. As healthcare systems across the globe increasingly seek innovative solutions to improve efficiency, enhance patient outcomes, and reduce operational costs, IoT emerges as a key enabler of this transformation. Despite the widely recognized benefits of IoT, its adoption in the healthcare sector, particularly within public hospitals in developing countries, remains limited and is still in the early stages. Therefore, understanding the factors influencing its adoption is essential for supporting effective adoption and advancing digital healthcare initiatives. This study aims to assess the validity and reliability of the instrument designed to identify the factors influencing the adoption of IoT technology in Jordanian public hospitals. A structured survey instrument collected a preliminary dataset from forty decision-makers in Jordanian public hospitals. The survey items were developed based on constructs derived from the Technology-Organizational-Environment (TOE) framework and the Human-Organization-Technology fit (HOT-fit) model, supported by relevant literature. Descriptive statistics were performed using SPSS, while the reliability and validity of the instrument were assessed using Partial Least Squares Structural Equation Modeling (PLS-SEM). The results demonstrated that the measurement instrument had acceptable levels of reliability and validity, confirming its suitability for use in the main study. This study enriches the existing research and enhances the broader understanding of IoT adoption in healthcare organizations, offering insights that can be useful to both practitioners and researchers in this field.

Author 1: Mutasem Zrekat

Author 2: Othman Bin Ibrahim

Keywords: Internet of Things; adoption; Jordan; TOE

Paper 18: Research on Network Flow Based on Statistical Analysis Methods

Abstract: To solve the problem of detecting unexpected situations in network traffic, it is proposed to determine the normal, unexpected states and behavior of the system using statistical methods. The statistical analysis methods used are the mean, variance, asymmetry coefficient, kurtosis coefficient, opposition coefficient and entropy coefficient. All statistical analysis methods allow for a deeper understanding of the uniqueness of the data. Each coefficient provides different measurement values, with which it is possible to determine the general characteristics of the data. The application of statistical analysis methods in network traffic is one of the most common methods for implementing the technology of detecting unexpected situations. For this, network traffic recorded in real time in laboratory conditions is used. Packet features of the network traffic are extracted from the recorded data, and then statistical analysis is performed using the extracted features. Markov chain is one of the most effective tools for analyzing situations and events occurring in network traffic. Markov chain is formed using the results of statistical analysis, and a Markov chain is constructed. This approach represents the state of the network flow in a probabilistic model and serves as an effective tool in monitoring the network flow.

Author 1: Maruf Juraev

Author 2: Inomjon Yarashov

Author 3: Adilbay Kudaybergenov

Author 4: Alimdzhan Babadzhanov

Author 5: Zilolaxon Mamatova

Keywords: Unexpected situation; mean; variance; asymmetry coefficient; kurtosis coefficient; opposition coefficient; entropy coefficient; packet analysis; markov chain; packet features

Paper 19: A Steel Surface Defect Detection Method Based on Lightweight Convolution Optimization

Abstract: Surface defect detection of steel, especially the recognition of multi-scale defects, has always been a major challenge in industrial manufacturing. Steel surfaces not only have defects of various sizes and shapes, which limit the accuracy of traditional image processing and detection methods in complex environments. However, traditional defect detection methods face issues of insufficient accuracy and high miss-detection rates when dealing with small target defects. To address this issue, this study proposes a detection framework based on deep learning, specifically YOLOv9s, combined with the C3Ghost module, SCConv module, and CARAFE upsampling operator, to improve detection accuracy and model performance. First, the SCConv module is used to reduce feature redundancy and optimize feature representation by reconstructing the spatial and channel dimensions. Second, the C3Ghost module is introduced to enhance the model’s feature extraction ability by reducing redundant computations and parameter volume, thereby improving model efficiency. Finally, the CARAFE upsampling operator, which can more finely reorganize feature maps in a content-aware manner, optimizes the upsampling process and ensures detailed restoration of high-resolution defect regions. Experimental results demonstrate that the proposed model achieves higher accuracy and robustness in steel surface defect detection tasks compared to other methods, effectively addressing defect detection problems.

Author 1: Cong Chen

Author 2: Ming Chen

Author 3: Hoileong Lee

Author 4: Yan Li

Author 5: Jiyang YU

Keywords: YOLOv9s; steel surface defect detection; C3Ghost module; SCConv module; CARAFE upsampling operator

Paper 20: Landslide Detection Method Based on Lightweight Convolution and Attention Mechanisms

Abstract: Landslide monitoring is a crucial component of geological disaster early warning systems. Traditional landslide detection methods often suffer from insufficient accuracy or low efficiency. To address these issues, this study proposes an improved landslide detection algorithm based on YOLOv11n, aiming to enhance both detection accuracy and efficiency by optimizing the model structure. First, the GhostConv module is introduced to reduce redundant computations, thereby improving computational efficiency. Additionally, the C3K2-SCConv optimization module is incorporated, which enhances feature extraction capability and improves the recognition of landslides at different scales by integrating multi-scale information and a weighted convolution strategy. Furthermore, the SimAM attention mechanism is implemented to adaptively adjust feature map weights, strengthening key features in landslide regions and improving detection accuracy. Experimental results demonstrate that the improved model achieves a mean average precision (mAP@0.5) of 83.3%, a precision of 85.5%, and a recall of 78.1%, representing increases of 2.0%, 3.2%, and 2.8%, respectively, compared to the baseline model. The proposed improvements provide a more accurate and efficient landslide detection method, contributing to the precision of geological disaster early warnings and enhancing the reliability of disaster prevention and mitigation efforts.

Author 1: Cong Chen

Author 2: Chengyang Zhang

Author 3: Ran Chen

Author 4: Jiyang YU

Keywords: YOLOv11n; landslide detection; GhostConv module; C3K2-SCConv module; SimAM attention mechanism

Paper 21: Integration of 2D-CNN and LSTM Networks for Enhanced Image Processing and Prediction in Alzheimer’s Disease

Abstract: The early diagnosis of Alzheimer’s disease remains a major challenge due to the complexity of magnetic resonance image interpretation and the limitations of existing diagnostic models. The slow memory loss associated with the gradual loss of thinking abilities, known as Alzheimer's disease, is the most common element of the illness. Effective early diagnosis is therefore essential to treatment; unfortunately, the traditional diagnostic procedure, which involves analyzing magnetic resonance images, is a complex process and prone to mistakes. This study aims to successfully merge these cognitive models with advanced deep learning techniques to enhance the diagnostic capabilities of Alzheimer’s disease using a fusion model with 3-dimensional convolutional neural networks and long short-term memory networks. The proposed approach uses three-dimensional convolutional neural networks to extract intricate features from volumetric magnetic resonance images, while long short-term memory networks analyze sequential data to identify key temporal patterns that indicate the progression of Alzheimer's disease. The dataset used in this study is the Alzheimer's Disease Neuroimaging Initiative dataset, which contains magnetic resonance images labeled into four categories: Non-Demented, Very Mild Demented, Mild Demented, and Moderate Demented. The dataset consists of 6,400 magnetic resonance images in total, split into training (70%), validation (15%), and testing (15%) sets. These outcomes demonstrate that the hybrid model improves predictive accuracy significantly over current benchmarks on this topic. This study highlights the importance of introducing deep learning models into clinical practice, thereby providing an efficient tool for early-stage Alzheimer’s disease diagnosis, ultimately improving patient outcomes through early and accurate intervention.

Author 1: Aya Mohamed Abd El-Hamed

Author 2: Mohamed Aborizka

Keywords: Alzheimer’s disease; magnetic resonance imaging; two-dimensional convolutional neural network; long short-term memory; deep learning; early detection

Paper 22: Audio-Visual Multimodal Deepfake Detection Leveraging Emotional Recognition

Abstract: Recently, there has been a significant reliance on the Internet. This creates a fertile environment for various risks, including fraud, privacy violations, and theft. The most common and dangerous risks at present are known as deepfakes. The development of deepfake technologies relies on advancements in artificial intelligence. Deepfake content can greatly affect privacy and security, posing a significant risk to many fields. Therefore, recent research has focused on mechanisms to detect real content from fake content. These mechanisms are classified into two main types: single-modal and multimodal detection. It is worth noting that the widespread deepfake technology has recently become more complex. This may hinder traditional single-mode detection methods in detecting video clips. In this study, we designed an effective multimodal fusion mechanism that integrates pre-trained audio, visual, and textual features. Our framework is based on three considerations: audio features, visual features, and emotion recognition. Emotion recognition focuses on three considerations: audio emotion, facial emotion, and sentiment of speech. We take advantage of the sentiment of speech to ensure there is consistency between audio and visual emotion with the meaning of words. As we achieved, the sentiment of speech makes our model more accurate and robust than when we used the audio-visual emotion inconsistency measures only. In our experiment, we used the FakeAVCeleb dataset, and we achieved 95.24% accuracy, proving our assumption of the impact of the sentiment of speech, the emotion of audio tone, and facial expressions to detect deepfakes.

Author 1: Alaa Alsaeedi

Author 2: Amal AlMansour

Author 3: Amani Jamal

Keywords: Machine learning; deepfake; multimodal; sentiment of speech; emotion recognition

Paper 23: A Novel Multi-Modal Deep Learning Approach for Real-Time Live Event Detection Using Video and Audio Signals

Abstract: Recent developments in live event detection have primarily focused on single-modal systems, where most applications are based on audio signals. Such methods normally rely on classification approaches involving the Mel-spectrogram. Single-modal systems, though effective in some applications, suffer from severe disadvantages in capturing the complexities of a real-world event, which thereby reduces their reliability in dynamically changing environments. This research study presents a novel multi-modal deep learning approach that combines audio and visual signals in order to enhance the accuracy and robustness of live event detection. The innovation lies in the use of two-stream LSTM pipelines, allowing for temporally consistent modeling of both input modalities while keeping a real-time processing pace through feature-level fusion. Unlike many of the recent transformer models, we are utilizing proven techniques (MFCC, 2D CNN, ResNet and LSTM) in a latency-aware and deployment-friendly architecture suitable for embedded and edge-level event detection. The AVE (Audio Video Events) dataset, consisting of 28 categories, has been used. For the visual modality, video frames undergo feature extraction through a 2D CNN ResNet and temporal analysis through an LSTM. Simultaneously, the audio modality employs MFCC (Mel Frequency Cepstral Coefficients) for feature extraction and LSTM to capture temporal dependencies. The features extracted from both audio and video modalities are concatenated for fusion. The proposed integration leverages the complementary nature of audio and visual inputs to create a more comprehensive framework. The outcome yields 85.19% accuracy in audio and video-based events due to the effective fusion of spatial and temporal cues from diverse modalities, outperforming single-modal baselines (audio-only or video-only models).

Author 1: Pavadareni R

Author 2: A. Prasina

Author 3: Samuthira Pandi V

Author 4: Ibrahim Mohammad Khrais

Author 5: Alok Jain

Author 6: Karthikeyan

Keywords: Multi-modality; feature fusion; early fusion; concatenation audio-video signals; convolutional neural network (CNN); Long Short-Term Memory (LSTM); Mel Frequency Cepstral Coefficients (MFCC); ResNet (Residual Network)

Paper 24: False News Recognition Model Based on Attention Mechanism and Multiple Features

Abstract: As the prevalence of social media continues to grow, the rapid and wide dissemination of false news has become a critical societal challenge, undermining public trust, creating social unrest, and distorting political discourse. Traditional fake news detection methods often rely solely on linguistic cues or shallow semantic analysis, which leads to limited accuracy and poor robustness, particularly when addressing emotionally biased or contextually complex content. To overcome these limitations, this study proposes a novel fake news recognition model based on a bidirectional gated recurrent unit combined with a self-attention mechanism, further enhanced by integrating sentiment polarity, textual metadata, and contextual semantic features. Experimental results show that the proposed model achieves a recognition accuracy of ninety-seven per cent and an F1 score of ninety-seven. In addition, it demonstrates the lowest mean absolute error, which is zero point one nine, and the shortest recognition time, requiring only zero point eight seconds after eighty iterations. The model also maintains over ninety-three per cent accuracy across news content with active, negative, and neutral emotional tones. The model offers a scalable and reliable framework for detecting false news, with strong adaptability to diverse content types and emotional expressions, thereby contributing to the advancement of automated misinformation identification in real-world applications.

Author 1: Qiongyao Suo

Author 2: Hongzhen Chang

Keywords: Fake news; attention mechanism; multiple features; bidirectional gated recurrent unit

Paper 25: A Reading-Aware Fusion Fact Reasoning Network for Explainable Fake News Detection

Abstract: The current growth of information exhibits an exponential trend, with fake news becoming a focal issue for both the public and governments. Existing fact-checking-based fake news detec-tion methods face two challenges: a heavy reliance on fact-checking reports, a lack of explanatory evidence related to the original reports, and a shallow level of feature interaction. To address these challenges, this study proposes a Reading-aware Fusion Fact Reasoning Network for explainable fake news de-tection. In the aspect of extractive evidence for explainability, a Hierarchical Encoding Layer is constructed to capture sen-tence-level and document-level feature representations, followed by a Fact Reasoning Layer to obtain report and sentence repre-sentations most relevant to the claim, thereby reducing the mod-el's reliance on fact-checking reports. Inspired by reading be-haviors, which often involve repeatedly reading the claim and corresponding report during information verification, the Read-ing-aware Fusion Layer is introduced to learn the deep interde-pendencies among the claim, evidence, and report feature repre-sentations, enhancing semantic integration. Extensive experi-ments were conducted on the publicly available RAWFC and LIAR fake news datasets. The experimental results demonstrate that RFFR outperforms leading advanced baselines on both datasets.

Author 1: Bofan Wang

Author 2: Shenwu Zhang

Keywords: Explainable fake news detection; fact reasoning; feature fusion

Paper 26: Solar-Net: Adaptive Fusion of Spatial-Temporal Features for Resilient Solar Power Generation Forecasting

Abstract: Solar power generation forecasting faces significant challenges due to intermittency and volatility, particularly under extreme weather conditions. This study proposes Solar-Net, a novel solar power generation prediction model based on a CNN+Transformer hybrid parallel architecture with an adaptive attention fusion mechanism. The CNN branch extracts spatial features from the power station layout and environmental conditions, while the Transformer branch models temporal dependencies in generation patterns. The core innovation lies in the adaptive attention fusion mechanism that dynamically adjusts branch weights according to real-time meteorological conditions, enabling the model to automatically adapt to varying environmental scenarios. Experiments were conducted on a comprehensive dataset containing over 50,000 observation points from two photovoltaic power stations. Results demonstrate that Solar-Net achieves superior performance compared to existing methods, with Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) improvements of 12.7% and 10.9%, respectively. Under extreme weather conditions such as dust storms, the model maintains prediction errors within 8.5% of peak power generation, representing a 45.7% average reduction compared to baseline methods. The multi-scale convolution design enhances prediction accuracy by 10.5% while reducing computational complexity by 21.3%. The proposed Solar-Net model provides a robust and efficient solution for solar power generation forecasting, demonstrating significant potential for improving grid dispatching efficiency and supporting renewable energy integration in power systems.

Author 1: Wenqian Su

Author 2: Jason See Toh Seong Kuan

Author 3: Xiangyu Shi

Author 4: Yuchen Zhang

Keywords: Solar power generation forecasting; hybrid deep learning; adaptive attention fusion; CNN+Transformer; extreme weather adaptability; sustainable development goal 7

Paper 27: Framework for Child Healthcare System Using Random Forest

Abstract: Proactive and customized approaches are necessary when it comes to the medical care of expectant mothers and children. Even if early and accurate disease prediction is based on readily available symptom information, it can significantly improve outcomes by promoting timely therapies. Extensive testing and specialist visits are common components of traditional diagnosis techniques, which may be costly and time-consuming, especially in situations with limited resources. This study reconnoiters the potential of using Random Forest, a powerful machine learning algorithm, to predict diseases in children and pregnant women based on the symptoms that they exhibit. This offers a possible choice for improved healthcare delivery and early risk assessment. Making predictions about childhood diseases, including pneumonia, malaria, and malnutrition, based on reported symptoms, can significantly lower morbidity and death. A Random Forest model can identify the probability of certain diseases and provide rapid referrals for additional testing and treatment when symptoms like fever, cough, dyspnoea, and weight loss are entered. Communities that are geographically remote and have limited access to specialized medical care should pay special attention to this. The early diagnosis of conditions including gestational diabetes, preeclampsia, and anemia during pregnancy is crucial for the mother's and the unborn child's health. Early detection of the ailment allows for the timely implementation of preventative measures, such as changing one's lifestyle or taking medication. The accuracy of the proposed Child healthcare system is 92% which is greater than other present methods. This analysis is based on the information provided by parents about the symptoms of their child’s diseases.

Author 1: Mahesh Ashok Mahant

Author 2: P. Vidyullatha

Keywords: Child healthcare system; random forest; machine learning; registration process; medicine

Paper 28: Enhancing Organizational Threat Profiling by Employing Deep Learning with Physical Security Systems and Human Behavior Analysis

Abstract: Organizations need a comprehensive threat profiling system that uses cybersecurity methods together with physical security methods because advanced cyber-threats have become more complex. The objective of this study is to implement deep learning models to boost organizational threat identification via human behavior assessment and continuous surveillance activities. Our method for human behavior analysis detects insider threats through assessments of user activities that include logon patterns along with device interactions and measurement of psychometric traits. CNN, together with Random Forest classifiers, has been utilized to identify behavioral patterns that indicate security threats from inside the organization. Our model uses labeled datasets of abnormal user behavior to properly differentiate between normal and dangerous user activities with high accuracy. The physical security component improves surveillance abilities through the use of MobileNetV2 for real-time anomaly detection in CCTV video data. The system receives training to detect security breaches and violent and unauthorized entry attempts, and specific security-related incidents. The combination of transfer learning and fine-tuning methodologies enables MobileNetV2 to deliver outstanding security anomaly detection alongside low power requirements, thus it fits into Security Operations Centers operations. Experiments using our framework operate on existing benchmark collection sets that assess cybersecurity, together with physical security threats. Experimental testing establishes high precision levels for detecting insider threats along with physical security violations by surpassing conventional rule-based methods. Security Operation Centers gain an effective modern threat profiling solution through the application of deep learning models. The investigation generates better organization defenses against cyber-physical threats using behavioral analytics together with intelligent surveillance systems.

Author 1: D. H. Senevirathna

Author 2: W. M. M. Gunasekara

Author 3: K. P. A. T. Gunawardhana

Author 4: M. F. F. Ashra

Author 5: Harinda Fernando

Author 6: Kavinga Yapa Abeywardena

Keywords: Deep learning; physical security; human behavior analysis; security operation centers; threat profiling

Paper 29: LaObese: A Serious Game Powered by Analytic Hierarchy Process for Culturally Tailored Childhood Obesity Prevention in Oman

Abstract: Childhood obesity is a growing public health concern in Oman, yet culturally appropriate digital tools for early prevention remain scarce. This study introduces LaObese, a mobile application and serious game designed to prevent obesity in Omani preschool children. The name LaObese derives from the Arabic word “La” (meaning “no”) and an abbreviation of “obesity”, reflecting the game’s mission to say, “no to obesity”. LaObese integrates gamification and behaviour modification strategies to encourage healthy lifestyle practices from an early age. Targeting children aged six, the initiative addresses the urgent issue of early childhood obesity, which is linked to long-term health complications. The development process incorporated a Multi-Criteria Decision-Making (MCDM) approach – specifically the Analytic Hierarchy Process (AHP) to identify and prioritise foods commonly consumed by children, ensuring the game’s nutritional content is both effective and culturally relevant. Food items were categorised (e.g., fruits, vegetables, proteins, beverages, desserts, and traditional Omani dishes) to align with preschool nutrition goals and local dietary habits. Findings from expert assessments (informed by national nutrition guidelines and local data from Salalah, Dhofar) highlight a growing preference for nutrient-rich traditional foods like dates and almonds among young children. LaObese is the first serious game in the Arab region to integrate an AHP-driven educational model for childhood obesity prevention. The platform facilitates collaboration between health professionals and educators in creating culturally tailored digital interventions, aiming to instil healthy eating habits in children through engaging gameplay.

Author 1: Nurul Akhmal Mohd Zulkefli

Author 2: Mukesh Madanan

Author 3: Zainab Mohammed Al-Nahdi

Author 4: Jayasree Radhamaniamma

Keywords: Serious games; childhood obesity; gamification; analytic hierarchy process (AHP); preschool nutrition

Paper 30: Anomaly Study of Computer Networks Based on Weighted Dynamic Network Representation Learning

Abstract: One of the foremost significant challenges in the continuously increasing technological environment is the requirement to secure the authenticity of data. Network security is a primary method for securing the confidentiality of data throughout communication, one of several types of data security assurance. To secure networks against additional cyberattacks, trustworthy Anomaly Detection (AD) is essential. The drawbacks of conventional AD are gradually increasing as various types of attacks and network changes continually evolve. The researchers of the present study propose a novel approach that incorporates Weighted Long Short-Term Memory (WLSTM) networks with Dynamic Network Representation Learning (DNRL) to address these problems, referring to it as the Weighted Dynamic Network Representation Learning (WDNRL) paradigm. This investigation develops the WLSTM utilizing the Weight of Evidence (WoE), which periodically determines weights to network features in the resulting network model. The WLSTM design functions as the network's coordinator, obtaining data from the recommended model, upgrading the representation, and aggregating the features. The findings showed that the proposed model achieved high accuracy rates of 99.85% for Denial of Service (DoS) attacks and 99.55% for Distributed Denial of Service (DDoS) attacks when evaluated using two datasets, NSL-KDD and CICIDS-2017, compared to different models. Additionally, the simulation's F1-scores, recall rates, and precision are all above average, indicating that it is capable of identifying many network anomalies with minimal false positives (FP).

Author 1: Xin Wei

Keywords: Network security; attacks; weighted dynamic network; anomaly detection; deep learning; LSTM

Paper 31: Deepfake Audio Detection Using Feature-Based and Deep Learning Approaches: ANN vs ResNet50

Abstract: The proliferation of algorithms and commercial tools for generating synthetic audio has sparked a surge in mis- information, especially on social media platforms. Consequently, significant attention has been devoted to detect such misleading content in recent years. However, effectively addressing this challenge remains elusive, given the increasing naturalness of fake audio. This study introduces a model designed to distinguish between natural and fake audio, employing a two-stage approach: an audio preparation phase involving raw audio manipulation, followed by modeling using two distinct models. The first model employed feature extraction through wavelet transformation, followed by classification using a machine learning Artificial Neural Network. The second model utilized ResNet50 architecture, a type of deep learning model, which resulted in improved accuracy. These findings underscore the effectiveness of deep learning approaches in audio classification tasks. Training data for the model is sourced from the DEEP-VOICE dataset, which comprises both genuine and synthetic audio generated by various deep-fake algorithms. The model’s performance is assessed using diverse metrics such as accuracy, F1 score, precision and recall. Results indicate successful classification of audio in 86% of cases. This research contributes to the field of Automatic Speech Recognition (ASR) by integrating advanced preprocessing techniques with robust model architectures to identify manipulated speech.

Author 1: Reham Mohamed Abdulhamied

Author 2: Sarah Naiem

Author 3: Mona M. Nasr

Author 4: Farid Ali Moussa

Keywords: Audio classification; automatic speech recognition; machine learning; deep learning; DEEP-VOICE

Paper 32: Fine-Tuning OpenAI GPT Chatbot in Western Saudi Dialect: A Case Study of Taibah University

Abstract: The current era is characterized by technological advancement and innovation, which affect various sectors. Numerous remarkable and alluring computer programs and applications have surfaced, including ones that aim to replicate human behavior. A chatbot is an example of an Artificial Intelligence (AI) computer program that uses natural language to mimic human conversations in voice or content. Even though a lack of Arabic chatbots, most of these chatbots use Modern Standard Arabic instead of Arabic dialects. This research presents the development and evaluation of a chatbot designed to respond to academic inquiries from university students using the Western Saudi dialect. A traditional Support Vector Machine (SVM) baseline model was first implemented to establish a reference point for performance. Subsequently, a fine-tuned version of Generative Pre-trained Transformer (GPT) 3.5-Turbo-0125 was developed using a culturally specific system prompt to enhance the model’s understanding of regional language and academic contexts. Evaluation was conducted through a multi-dimensional framework combining human assessments, BERTScore semantic similarity measurements, and GPT-4-based automatic judging. With human assessors determining that 85% of GPT-3.5's replies to 132 messages of test data were appropriate, the transformer-based model clearly outperformed the SVM baseline, which had an accuracy of 42.86% on 20 messages of test data. These findings highlight the importance of cultural and contextual fine-tuning in building effective conversational agents for dialectal Arabic communities. The research contributes to the growing field of localized AI by demonstrating how advanced language models can be adapted to serve specialized linguistic and academic needs.

Author 1: Maimounah Alhujaili

Author 2: Ruqayya Abdulrahman

Keywords: Artificial intelligence (AI); large language model (LLM); generative pre-trained transformer (GPT); Modern Standard Arabic (MSA); Western Saudi dialect

Paper 33: Automatic Detection of Natural Disasters Using Faster R-CNN with ResNet50 Backbone

Abstract: Natural disasters pose significant threats to human life and infrastructure. Timely detection and assessment of these events are crucial for effective disaster management. This study proposes an automatic detection system for natural disasters using aerial imagery. Accurate and timely detection of natural disasters is critical for minimizing their impact and supporting emergency response efforts. This study presents a comparative analysis of deep learning architectures for natural disaster detection using satellite and aerial imagery. Four models were evaluated as baseline CNN, ResNet50, Faster-CNN, and Faster R-CNN with a ResNet50 backbone using standard classification metrics. The results demonstrate that deeper and more sophisticated models significantly enhance detection performance. While the baseline CNN achieved modest results with 85.3% accuracy, integrating residual learning in ResNet50 improved accuracy to 92.7%. Region-based models further boosted performance, with Faster-CNN and Faster R-CNN attaining 95.1% and 97.1% accuracy, respectively. The superior performance of the Faster R-CNN with ResNet50 highlights its robustness and suitability for real-time disaster monitoring, offering a scalable and reliable solution for operational deployment in disaster management systems.

Author 1: Shereen Essam Elbohy

Author 2: Mona M. Nasr

Author 3: Farid Ali Mousa

Keywords: Natural disasters detection; satellite imagery; convolutional neural networks (CNN); transformers; deep learning; ResNet50; proactive monitoring; faster R-CNN; disaster prevention; computer vision

Paper 34: Fusion of CNN and Transformer Architectures for Proactive Wildfire Detection in Satellite Imagery

Abstract: Wildfires pose a significant threat to ecosystems, human settlements, and air quality, necessitating advanced detection and mitigation strategies. Traditional wildfire detection methods often rely on manual observation and conventional machine learning approaches, which may lack efficiency and accuracy. This study proposes a novel deep learning model based on the ConvNeXt-Small architecture, a hybrid design that fuses the strengths of Convolutional Neural Networks (CNNs) and Transformer-inspired mechanisms, enabling more comprehensive analysis of wildfire patterns in satellite imagery. The model was trained using the Adam optimizer, which provides efficient convergence and adaptive learning. The dataset used consists of real-world satellite images collected from wildfire-affected regions in Canada, covering various geographic and seasonal conditions to reflect real environmental diversity. The results underscore the potential of ConvNeXt-based architecture for real-time, high-precision wildfire detection, offering a powerful tool for early intervention, disaster mitigation, and environmental monitoring efforts.

Author 1: Shereen Essam Elbohy

Author 2: Mona M. Nasr

Author 3: Farid Ali Mousa

Keywords: Wildfire detection; satellite imagery; convolutional neural networks (CNN); transformers; deep learning; hybrid model; proactive monitoring; remote sensing; disaster prevention; computer vision

Paper 35: Analyzing the Impact of Robotic Process Automation (RPA) on Productivity and Firm Performance in the Service Sector

Abstract: Robotic Process Automation (RPA) has emerged as a transformative technology in the service sector, enabling organizations to automate repetitive and rule-based tasks with minimal human intervention. This study investigates the impact of RPA implementation on productivity and overall firm performance within service-oriented businesses. Using a mixed-method approach, quantitative data were collected from 50 service firms that have adopted RPA technologies, complemented by qualitative insights from managerial interviews. The findings reveal that RPA significantly enhances operational efficiency by reducing process cycle times, minimizing errors, and lowering operational costs. These productivity gains directly contribute to improved financial outcomes and customer satisfaction, key indicators of firm performance. Furthermore, the study highlights critical success factors such as employee training, change management, and technology integration that influence the effectiveness of RPA deployment. However, challenges related to workforce adaptation and initial investment costs are also discussed. This research provides valuable empirical evidence for service sector firms considering RPA adoption, emphasizing that strategic implementation can lead to sustainable competitive advantages. The study contributes to the growing body of knowledge on digital transformation by linking RPA technology with measurable improvements in productivity and firm performance, offering practical recommendations for managers and policymakers aiming to optimize automation strategies.

Author 1: Miftakul Huda

Author 2: Agus Rahayu

Author 3: Chairul Furqon

Author 4: Mokh Adib Sultan

Author 5: Neng Susi Susilawati Sugiana

Keywords: Robotic process automation; productivity improvement; firm performance; service sector; digital transformation; operational efficiency

Paper 36: Systematic Literature Review on Artificial Intelligence-Driven Personalized Learning

Abstract: Artificial Intelligence (AI) is widely used in various contexts, including education at different levels, such as K-12 (kindergarten through 12th grade) and higher learning. The impact of AI in education is becoming increasingly significant, making the academic sphere more effective, personalized, global, context-intensive, and asynchronous. Despite the publication of several systematic literature reviews, mapping studies, and reviews on the use of AI in education, there is still a lack of reviews focusing on personalized learning (PL) frameworks, models, and approaches at various levels especially the pre-university level for Science, Technology, Engineering, and Mathematics (STEM) subjects. To address this gap, our work presents a systematic literature review of AI-driven PL models, frameworks, and approaches published over the past ten years from 2013 to 2023, extracted from the Scopus database. This review focuses on the AI techniques used, personalized learning elements, components, attributes, and the possibility of replicating the technique in pre-university level studies, and gaps or prospects that will attract further research. The study reviewed 69 articles, downloaded via the Scopus database, and reported the most used AI techniques, PL components or factors, trends, and prospects for future research. The results show that most existing studies focus on higher learning that requires further research at the pre-university level. In addition, machine learning and deep learning are identified as the most suitable and frequent techniques besides other technologies, knowledge delivery, learners’ needs, behavior and interest as the most required components for personalized systems in diverse fields. In terms of publication output by country, the study indicates that Switzerland, USA, UK, and China are leading contributors to PL research. Thus, this study calls for further research on AI-driven personalized learning that thoughtfully integrates educational theories, subject-specific content, and industry needs to enhance outcomes and learner satisfaction.

Author 1: Anas Usman Inuwa

Author 2: Shahida Sulaiman

Author 3: Ruhaidah Samsudin

Keywords: Personalized learning; model; framework; approach; technique; systematic literature review; personalized learning components; artificial intelligence

Paper 37: Mobile Application Using Convolutional Neural Networks for Preliminary Diagnosis of Rosacea

Abstract: Rosacea is a chronic skin disease affecting millions of people worldwide, characterized by redness and inflammatory lesions on the face. Given the need to improve early detection, this research aims to develop a mobile application using convolutional neural networks to improve the preliminary diagnosis of rosacea. For this purpose, increases in sensitivity, specificity and accuracy percentages were evaluated. The study was applied, with a quantitative approach and an experimental design, specifically pre-experimental. The study variable was the preliminary diagnosis of rosacea, and the sample consisted of 100 images: 50 from rosacea patients and 50 from healthy people. The technique used for data collection was observation. The results of the implementation of the mobile application showed an increase in sensitivity of 2.7%, specificity of 1.97% and accuracy of 0.10%. In conclusion, the use of the mobile application with convolutional neural networks improves the preliminary diagnosis of rosacea by optimizing the indicators evaluated.

Author 1: Angie Fiorella Sapaico-Alberto

Author 2: Rosalynn Ornella Flores-Castañeda

Keywords: Convolutional neural networks; mobile application; rosacea; preliminary diagnosis; sensitivity; specificity

Paper 38: Hybrid PSO-ACO Optimization for Rice Leaf Disease Classification Using Random Forest and Support Vector Machines

Abstract: This study proposes a hybrid machine learning framework for rice leaf disease detection by combining handcrafted feature extraction with metaheuristic optimization and classical classifiers. Using a dataset of 6,000 rice leaf images across seven classes, features including color, texture, shape, and edge were extracted and optimized using Spider Monkey Optimization (SMO), Particle Swarm Optimization (PSO), and Ant Colony Optimization (ACO). Classification was conducted using Random Forest Classifier (RFC) and Support Vector Classifier (SVC), both with and without hyperparameter tuning. Experimental results revealed that PSO consistently outperformed other optimizers, achieving 91.00% accuracy with RFC and 94.64% with SVC when all features and optimal parameters were used. While SMO also showed strong performance, ACO yielded less consistent results. These findings highlight the importance of combining comprehensive feature engineering with adaptive optimization strategies to improve classification accuracy. Compared to previous SMO-based approaches, the proposed PSO-ACO framework demonstrated improved stability and scalability. The proposed framework is interpretable, efficient, and scalable, making it suitable for practical deployment in precision agriculture. Future research directions include integrating deep learning with handcrafted features, developing adaptive metaheuristics, and implementing real-time mobile detection systems.

Author 1: Avip Kurniawan

Author 2: Tri Retnaningsih Soeprobowati

Author 3: Budi Warsito

Keywords: Rice leaf disease; particle swarm optimization (PSO); support vector machine (SVM); feature extraction; precision agriculture

Paper 39: Reinforcement Learning for Real-Time Scheduling in Dynamic Reconfigurable Manufacturing Systems

Abstract: This study presents a novel application of Reinforcement Learning (RL) and Deep Reinforcement Learning (DRL) for scheduling optimization in Reconfigurable Manufacturing Systems (RMFS). The performance of these approaches is quantitatively evaluated and compared with traditional scheduling methods, specifically Shortest Processing Time (SPT) and Earliest Due Date (EDD), across several key metrics, including makespan, tardiness, resource utilization, and adaptability to disturbances. Our results show a significant reduction in makespan, with RL achieving a 20% improvement and DRL a 28.57% improvement over SPT. Moreover, RL and DRL outperform classical methods in minimizing tardiness and improving resource utilization. DRL also demonstrates superior adaptability under dynamic disruptions such as machine breakdowns, with only a 5% deviation in makespan compared to 16.67% for SPT. These findings confirm the benefits of RL and DRL for real-time decision-making in dynamic manufacturing environments. The study discusses the robustness and scalability of RL and DRL approaches, as well as the challenges related to their computational cost. The novelty lies in integrating RL and DRL into RMFS scheduling to offer a scalable, adaptive solution that improves production efficiency.

Author 1: Salah Hammedi

Author 2: Abdallah Namoun

Author 3: Mohamed Shili

Keywords: Adaptability; deep reinforcement learning (DRL); makespan; manufacturing systems; reinforcement learning (RL); resource utilization; scheduling optimization; shortest processing time (SPT); tardiness; traditional scheduling methods

Paper 40: A Proposed Framework for Loan Default Prediction Using Machine Learning Techniques

Abstract: The accurate prediction of loan defaults is critical for the risk management strategies of financial institutions. Traditional credit assessment approaches have often relied on subjective judgment, leading to inconsistent decisions and heightened financial risk. This study investigates the application of machine learning techniques—namely Random Forest, Decision Tree, and Gradient Boosting—to predict loan defaults using customer data from the Agricultural Bank of Egypt. The research emphasizes the role of feature selection in enhancing model performance, utilizing both embedded and recursive methods to isolate key predictive attributes. Among the evaluated features, loan balance, due amount, and delinquency history emerged as the most influential, while demographic variables like gender and employment status were found to be less significant. The Decision Tree model demonstrated superior performance with an overall accuracy of 88%, a recall of 53%, and a specificity of 89%, making it the most effective among the tested classifiers. The findings highlight the importance of combining robust feature selection with interpretable models to support informed decision-making in banking.

Author 1: Mona Aly SharafEldin

Author 2: Amira M. Idrees

Author 3: Shimaa Ouf

Keywords: Random forest; decision trees; gradient boosting machines; feature selection; feature importance; loan default

Paper 41: Advanced AI-Driven Safety Compliance Monitoring in Dynamic Construction Environment

Abstract: Construction safety is a critical global concern due to the high-risk environment faced by workers, with accidents often leading to serious injuries and fatalities. To enhance construction management, this study proposes a scalable deep-learning model for real-time compliance monitoring of safety regulations. The research gap addressed is the lack of real-time, scalable AI solutions for safety compliance monitoring in dynamic construction environments. The YOLOv11n model was trained and evaluated to identify and track the use of safety helmets and vests in extreme dynamic environments, ensuring timely detection of non-compliance. It is hypothesized that the YOLOv11n model will outperform baseline models in accuracy and real-time monitoring speed. The YOLOv11n model outperformed other baseline models, with precision, recall, and mean average precision scores of 89.5%, 85%, and 91.6%, respectively, and a real-time processing speed of 71.68 FPS. Its lightweight size and performance make it suitable for deployment. Integrated with a person-detection framework, the system provides real-time desktop alerts for safety violations, enhancing safety compliance. These findings contribute to construction automation by advancing scalable AI-driven solutions for proactive safety compliance, reducing accidents, and improving operational efficiency on construction sites.

Author 1: Aisha Hassan

Author 2: Ali H. Hassan

Author 3: Yasmin Christensen

Author 4: Hussain Alsadiq

Keywords: YOLOv11n; personal protection equipment (PPE); construction safety; real-time object detection; deep learning; AI-driving compliance systems

Paper 42: Reinforcement Learning Improves SVM-Driven Algorithms for Classifying Multi-Sensor Data for Medical Monitoring

Abstract: Multi-sensor data in medical monitoring includes waveform changes in physiological signals and time-series characteristics of disease progression. These features typically exhibit high-dimensionality, large-scale, and time-varying characteristics. Nonlinear relationships exist between these features, increasing the difficulty of data processing and feature extraction, thereby reducing the classification capabilities of related algorithms. This study proposes a multi-sensor data classification processing method in medical monitoring based on reinforcement learning improved SVM. The algorithm employs the DBSCAN algorithm combined with Euclidean distance for clustering and data collection of multi-sensor data. Discrete wavelet transform is used to remove interference noise from the data, followed by convolutional neural networks for signal feature extraction from the denoised data. The Q-learning algorithm in reinforcement learning is used to improve the traditional SVM, with the extracted signal features input into the improved SVM. The classification results of medical monitoring multi-sensor data are output via a regression function. The experimental results show that the denoising results of medical monitoring data of the method are high, the signal-to-noise ratio is high, and the Kappa coefficient reaches up to 0.98. Therefore, it shows that the method can accurately classify medical monitoring multi-sensor data.

Author 1: Zhiwei Xuan

Author 2: Yajie Liu

Keywords: Reinforcement learning; improved SVM; medical monitoring; multi-sensor; data; classification processing

Paper 43: Developing an Ontology-Driven and Governance-Integrated Method for Information Dashboard Design

Abstract: Despite the increasing reliance on information dashboards across industries, dashboard design practices remain fragmented, lacking standardized methodologies, ontological formalization, and governance integration. Addressing these gaps, this study develops a method to guide dashboard design by embedding ontological modeling and Information Governance (IG) principles. Two complementary artifacts are proposed: the Information Dashboard Design Ontology (IDDO) and the Information Dashboard Design Method (IDDM) Canvas. Using Design Science Research Methodology (DSRM) and a Unified Ontological Approach (UOA), IDDO formalizes tacit dashboard design knowledge into a structured framework, while the IDDM Canvas operationalizes this ontology into a practical design tool. Validation through the Ontological Unified Modeling Language (OntoUML) Plugin and conceptual assessment based on Unified Foundational Ontology (UFO) principles confirmed internal consistency and ontological soundness. The resulting framework integrates twelve dashboard design building blocks with eight IG principles to ensure rigor and governance alignment. The application of the IDDM Canvas demonstrated its utility in facilitating structured, replicable dashboard development. While the evaluation focused primarily on conceptual validation, future studies are recommended to empirically assess the framework’s practical effectiveness across various domains and real-world projects.

Author 1: Ahadi Haji Mohd Nasir

Author 2: Nik Habibullah Nik Mohd Nizam

Author 3: Mohd Khairul Maswan Mohd Redzuan

Author 4: Mohammad Nazir Ahmad

Keywords: Information dashboard; ontological modelling; information dashboard design ontology (IDDO); information dashboard design method (IDDM) canvas; information governance (IG); unified ontological approach (UOA); design science research methodology (DSRM)

Paper 44: Machine Learning and 5G Edge Computing for Intelligent Traffic Management

Abstract: The integration of fifth-generation (5G) communication technology and Artificial Intelligence (AI) is reshaping urban mobility by enabling intelligent transportation systems and smarter cities. This synergy allows real-time traffic management, predictive maintenance, and enhanced autonomous driving, supported by high-speed, low-latency networks and advanced data analytics. By leveraging 5G’s strong connectivity, AI systems can process massive datasets to address urban challenges such as traffic congestion, environmental sustainability, and public safety. This study presents a framework that combines 5G and AI to optimize traffic management through dynamic congestion prediction and real-time routing, supported by edge computing. It highlights the benefits of improving traffic flow, reducing emissions, and enhancing overall urban mobility efficiency. In addition, it discusses key challenges including data privacy concerns, cybersecurity risks, and the high cost of infrastructure deployment. By analyzing existing technologies and proposing an AI-driven, 5G-enabled system model, this study aims to bridge the gap between theoretical advancements and practical urban implementations. The findings provide insights into scalable, efficient solutions for the future of smart transportation networks and offer directions for further research in this dynamic and evolving field.

Author 1: Talbi Chaymae

Author 2: Rahmouni Mhamed

Author 3: Ziti Soumia

Keywords: 5G Edge computing; traffic management; dynamic routing; smart cities; machine learning

Paper 45: A Rule-Based Framework for Clothing Fit Recommendation from 3D Body Reconstruction

Abstract: This research presents a comprehensive framework for body size estimation that accurately derives anthropometric measurements—specifically, the circumferences of the waist and hips—from a singular image by utilizing OpenPose for joint localization and SMPLify-X for precise 3D body modeling. The proposed methodology involves projecting the generated three-dimensional model onto a horizontal plane and applying a convex hull geometric assessment to extract relevant body measurements. These derived measurements are then classified into standardized clothing size predictions (XS–XL) via a transparent rule-based classification system suitable for e-commerce sizing and virtual fitting applications. Empirical validation conducted on the Agora dataset substantiated the framework's reliability across diverse body types, demonstrating strong consistency with industry sizing standards. The method is non-intrusive and interpretable, effectively addressing practical challenges in automated human pose estimation for retail contexts. Limitations include constraints related to body posture and potential clothing interference; however, the modular design enables enhancements such as integrating chest circumference measurements and mobile deployment. This scholarly contribution thus provides a robust, accessible solution for automated, image-based clothing size recommendations.

Author 1: Hamid Ouhnni

Author 2: Acim Btissam

Author 3: Belhiah Meryam

Author 4: Benachir Rigalma

Author 5: Soumia Zit

Keywords: Body size estimation; SMPLify-X; OpenPose; 3D body modeling; clothing size prediction; e-commerce sizing; human pose estimation

Paper 46: TL-MC-ShuffleNetV2: A Lightweight and Transferable Framework for Elevator Guideway Fault Diagnosis

Abstract: This study presents TL-MC-ShuffleNetV2, a lightweight and transferable fault diagnosis framework designed for elevator guideway vibration analysis. To tackle challenges such as limited labeled data and the constraints of real-time deployment, the approach integrates Variational Mode Decomposition (VMD) for multi-scale signal separation and employs a customized 1D ShuffleNetV2 backbone with multi-channel (MC) inputs. Squeeze-and-Excitation (SE) attention modules are embedded throughout the network to enhance channel-wise feature sensitivity. A transfer learning (TL) strategy is adopted, in which the model is initially trained using the Case Western Reserve University (CWRU) bearing dataset and subsequently adapted to the elevator domain by freezing early convolutional layers while fine-tuning higher-level layers. Evaluation results demonstrate that the proposed framework achieves a classification accuracy of 96.4%, alongside significantly reduced inference time and parameter complexity. Comparative and ablation experiments further validate the individual contributions of VMD preprocessing, SE modules, and transfer learning to model performance. Overall, the method exhibits strong adaptability, computational efficiency, and suitability for deployment in smart elevator monitoring systems under Industry 4.0 environments.

Author 1: Zhiwei Zhou

Author 2: Xianghong Deng

Author 3: Xuwen Zheng

Author 4: Chonlatee Photong

Keywords: Transfer learning; elevator guideway; vibration signal analysis; fault diagnosis; lightweight deep neural network; squeeze-and-excitation attention; smart maintenance

Paper 47: AI-Driven Intrusion Detection Systems for Securing IoT Healthcare Networks

Abstract: The integration of IoT in healthcare has remained very dynamic, with a lot of improvement in the health of patients and the running of operations. Integration also comes with new risks and threats, raising IoT healthcare networks as cyber victims with great potential. This study explores an AI-based solution to defend healthcare IoT networks against intrusions. Therefore, using the most superior machine learning algorithms and deep learning expertise, it is concluded that a credible IDS would be built eventually to be able to detect and neutralize security threats in a live environment. The proposed IDS are trained and tested on a large, rich data set of IoT healthcare security incidents and features like CNN and RNN. Our system has learned to identify numerous and different types of cyber threats, such as Malware, Ransomware, Unauthorized access, data breaches, and many more, with better accuracy and even fewer false positives. This study proves that IDS backed by Artificial Intelligence is effective in improving the security status of IoT healthcare networks, organization’s control over crucial patient information, and thus, the maintenance of the continuous provision of healthcare services.

Author 1: Muhammad Sajid Nawaz

Author 2: Muhammad Ahsan Raza

Author 3: Binish Raza

Author 4: Manal Ahmad

Author 5: Farial Syed

Keywords: IoT; intrusion detection system (IDS); convolutional neural network (CNN); recurrent neural network (RNN); cybersecurity

Paper 48: Leveraged Cognitive Data Analytics and Artificial Intelligence to Enhance Sustainable Agricultural Practices: A Systematic Review

Abstract: This systematic review examines the transformative role of Cognitive Data Analytics (CDA) and Artificial Intelligence (AI) in advancing sustainable agricultural practices, with a primary objective to evaluate their applications in Precision Agriculture (PA), Internet of Things (IoT), smart irrigation, and Geographic Information Systems (GIS) from 2020 to 2025. Key findings highlight AI predictive modeling, IoT real-time monitoring, and GIS spatial analysis improving crop yields, water conservation, and environmental management. Challenges such as high costs, technical expertise gaps, and regional disparities hinder adoption. The review underscores the need for supportive policies and farmer training to enhance food security and sustainability by 2030.

Author 1: Wongpanya S. Nuankaew

Author 2: Patchara Nasa-Ngium

Author 3: Pratya Nuankaew

Keywords: Cognitive data analytics; sustainable agriculture; harnessing AI for agriculture; precision agriculture; environmental sustainability; food security

Paper 49: A Novel Approach for Enhancing Advanced Encryption Standard Performance and Cryptographic Resilience

Abstract: Advanced Encryption Standard (AES) encrypts data in blocks of sixteen bytes to secure confidential data stored in the cloud. For cloud-based systems, enhancements in existing encryption techniques are necessary as the nature of cyber threats evolves and computational speed becomes increasingly critical. This study presents an enhanced design of AES that substitutes two special operations, Byte Transformation and Bits Permuted Bytes, with the conventional S-Box operation to improve the speed and security of the encryption method. The suggested round structure in the new approach of AES, which preserves the original data block size, consists of the following operations: Byte Transformation, Shift Rows, Bits Permuted Bytes, Add Round Key, and Mix Columns. The analysis of the strict avalanche effect, correlation coefficient, entropy, execution time, and throughput outcomes confirms that the developed scheme improves the security and processing speed.

Author 1: Muthu Meenakshi Ganesan

Author 2: Sabeen Selvaraj

Keywords: Cryptography; NIST; AES; block cipher; key expansion; symmetric encryption; galois field; statistical techniques; cryptanalysis

Paper 50: An Enhanced LSTM Model Based on Feature Attention Mechanism and Emotional Intelligence for Advanced Sentiment Analysis

Abstract: Sentiment analysis, a crucial yet complex task in natural language processing (NLP), is extensively employed to identify sentiment polarity within user-generated content. Traditional deep learning methods for textual sentiment analysis often overlook the influence of emotional modulation on extracting sentiment features. At the same time, their attention mechanisms primarily operate at the word or sentence levels. Such oversight of higher-level abstractions may hinder the learning of nuanced sentiment patterns, ultimately damaging the accuracy of sentiment analysis. Addressing these gaps, this study proposed a novel framework, the Two-State Enhanced LSTM (TS-ELSTM), which integrated Emotional Intelligence (EI) and a Feature Attention Mechanism (FAM) to enhance the identification of relevant features during selection. Furthermore, this study employed a dual-phase training strategy of LSTM to accelerate learning and minimize information loss. A dynamic topic-level attention mechanism is also introduced to optimize hidden text representation weights. By integrating EI with a topic-level attention mechanism, the proposed framework efficiently extracts valuable features and enhances the feature learning ability of the conventional LSTM model. This novelty attains emotion-aware learning through two key components: an emotion modulator and an emotion estimator, which successfully normalize the system’s learning dynamics by combining emotional context. The experimental outcomes demonstrated that the proposed approach achieved an accuracy of 84.20%, 94.12% using MR and IMDB, respectively. The proposed approach significantly improves sentiment analysis accuracy, outperforming traditional deep learning models by a notable margin.

Author 1: Muhammad Naeem Aftab

Author 2: Dost Muhammad Khan

Author 3: Muhammad Zulqarnain

Author 4: Muhammad Rizwan Akram

Keywords: Sentiment analysis; emotional intelligence; attention mechanism; two-state LSTM; long-term dependencies

Paper 51: Image Quality Assessment Based on Feature Fusion and Local Adaptation

Abstract: No-reference image quality assessment (NR-IQA) aims to evaluate the perceptual quality of images without access to corresponding reference images and has broad applications in real-world image processing scenarios. However, existing NR-IQA methods often suffer from limited accuracy and generalization, especially under complex and diverse distortion types. To address this challenge, we propose Inc-LAENet, a novel NR-IQA framework that leverages multi-scale deep residual representations, integrates feature fusion mechanisms, and incorporates a local adaptive perception module to achieve improved assessment accuracy and generalization. Specifically, ResNet50 is employed to extract hierarchical residual features, an enhanced Inception-style module (Inc-s) strengthens sensitivity to various distortion patterns, and a lightweight local adaptive extraction module efficiently captures fine-grained structural information. Extensive experiments demonstrate the effectiveness of the proposed method, achieving SROCC values of 0.967 and 0.935 on the synthetic distortion datasets LIVE and CSIQ, and 0.852 and 0.898 on the authentic distortion datasets LIVEC and KonIQ-10k, respectively. These results confirm that Inc-LAENet provides a robust and efficient solution for NR-IQA tasks across both synthetic and real-world scenarios.

Author 1: Minjuan GAO

Author 2: Yankang LI

Author 3: Xuande ZHANG

Keywords: No-reference image quality assessment; deep learning; multi-scale; feature fusion; local adaptation

Paper 52: Advancing Aerodynamic Coefficient Prediction: A Hybrid Model Integrating Deep Learning and Optimization Techniques

Abstract: The aerospace industry increasingly relies on predictive models for aerodynamic coefficients to enhance design, performance, and optimization. While traditional methods like Computational Fluid Dynamics (CFD) and wind tunnel simulations offer accurate predictions, they are computationally intensive and time-consuming. This study explores a novel approach that fuses advanced Deep Learning (DL) architectures with Optimization Techniques to achieve faster and more accurate predictions of aerodynamic coefficients. Building on the foundation of Convolutional Neural Networks (CNNs), we introduce hybrid models that integrate Evolutionary Algorithms and Gradient-Based Optimization to improve the accuracy, generalization, and adaptability of predictions. The proposed framework is validated on datasets derived from CFD simulations and wind tunnel experiments, demonstrating superior accuracy, reduced computational cost, and robust performance across diverse aerodynamic conditions. This study highlights the potential of combining DL and optimization methods as a transformative tool for real-time aerodynamic analysis, paving the way for more efficient Aerospace Design and decision-making. Future research directions include expanding the model to handle complex geometries and dynamic flight conditions.

Author 1: Jad Zerouaoui

Author 2: Rachid Ed-daoudi

Author 3: Badia Ettaki

Author 4: El Mahjoub Chakir

Keywords: Aerodynamic coefficients; computational fluid dynamics; deep learning; convolutional neural networks; optimization techniques; evolutionary algorithms; gradient-based optimization; aerospace design

Paper 53: AI-Powered Assessment of Resistance to Change in the Context of Digital Transformation

Abstract: Digital transformation is a key driver of business evolution, but it comes with significant challenges, particularly employee resistance to change. This resistance can manifest in various forms, ranging from explicit opposition to more subtle hesitation toward new practices. Its underlying causes are diverse, including fear of the unknown, loss of control, and dissatisfaction with perceived transformations. Understanding employee perceptions is, therefore, crucial to adapting digital initiatives and ensuring successful adoption. However, existing methods for assessing resistance, which rely on closed-ended questionnaires and binary classifications, have limitations. They restrict the expression of opinions and fail to provide a nuanced segmentation of employees’ stances toward change. In this context, this study proposes an innovative and automated methodology that combines specialized zero-shot LLMs and prompt engineering techniques to analyze resistance to change. It is based on the allies strategy, a concept derived from sociodynamic theory and widely applied in change management, which seeks to more precisely differentiate employee attitudes based on their level of synergy or antagonism toward a new project or transformation initiative. To evaluate the effectiveness of the proposed approach, an experiment was conducted on an annotated dataset comprising a hundred employee responses. Two prompt engineering strategies were explored and applied to six zero-shot models to assess their ability to accurately classify expressed attitudes. The findings underscored, on one side, the significance of prompt structuring in enhancing classification efficacy and, on the other side, the preeminence of DeBERTa-v3-large-zeroshot, which demonstrated itself as the most exemplary model, even exceeding GPT-4, one of the most sophisticated and cutting-edge language models currently accessible.

Author 1: Bachira Abou El Karam

Author 2: Tarik Fissaa

Author 3: Rabia Marghoubi

Keywords: Resistance to change; digital transformation; zero-shot LLMs; prompt engineering; allies strategy

Paper 54: Self-Supervised Method for Risky Situation Detection in Road Traffic Sequences Using Video Masked Autoencoder

Abstract: Road traffic accidents are a significant public health issue, particularly in developing nations, where infrastructure and traffic monitoring systems may be limited. Risky situations including sudden stopping, lane switching, and near-misses can lead to accidents. In this study, we present an original approach for recognizing risky situations in road traffic sequences using Video Masked Autoencoder (VideoMAE), a self-supervised deep learning model built upon Vision Transformer architecture. By applying a pre-trained VideoMAE on a large dataset of videos and fine-tuning it on labeled traffic sequences categorized as risky or non-risky, our model learns spatiotemporal features without requiring extensive manual labeling. The method achieves high accuracy on testing data, demonstrating strong potential for high-risk detection with an accuracy of 95%. This study highlights the promise of self-supervised video representation learning for real-world safety applications and paves the way for the development of intelligent traffic monitoring and crash prevention tools.

Author 1: Abdelhafid Berroukham

Author 2: Mohammed Lahraichi

Author 3: Khalid Housni

Keywords: Video processing; risk detection; VideoMAE; vision transformer; deep learning; computer vision

Paper 55: Vision-Based Vehicle Classification Using Deep Learning Model

Abstract: Vehicle classification offers intelligent solutions for road traffic monitoring by enabling future prediction planning and decision making. Predictive analytics can be used to predict traffic congestion based on the types of vehicles on the road. In this research, the reliability of deep learning based models for vision-based vehicle classification is investigated. Four models of You Only Look Once (YOLO) are investigated, namely YOLOv5s, YOLOv5x, YOLOv10n, and YOLOv12n. These models were trained and evaluated on a vehicle dataset comprising five vehicle classes, which are Ambulance, Bus, Car, Motorcycle, and Truck, with a total number of 1103 images. From the experiment conducted, YOLOv10n achieved the highest performance measure of mAP@0.5 with 0.859 across all vehicle classes, including per-class evaluation, demonstrating superior detection compared to the other models. Finally, the results indicate that the YOLOv10n model can be used in vision-based vehicle classification.

Author 1: Ahsiah Ismail

Author 2: Amelia Ritahani Ismail

Author 3: Muhammad Afiq Mohd Ara

Author 4: Asmarani Ahmad Puzi

Author 5: Suryanti Awang

Keywords: YOLO; vehicle classification; deep learning; traffic monitoring

Paper 56: Comparative Analysis of Rank and Roulette Wheel Selection Strategies in Genetic Algorithms for Spatial Layout Optimization

Abstract: Autonomous urban planning, facility layout design, and interior design are critical and meticulous tasks that require the optimization of space arrangement. One of the main purposes of space arrangement is to achieve high space utilization with a non-complex arrangement for emergency assistance, particularly to enhance pedestrian safety in panic situations. This study explores the optimization of spatial layouts by employing Genetic Algorithms (GA) due to their robust search capabilities. However, spatial layout size limitations may affect the search capability and significantly impact space arrangement and utilization. Hence, this study presents a comparative study of two GA selection operator methods: Rank Selection (RS) and Roulette Wheel Selection (RWS) for determining the effectiveness in optimizing spatial layout arrangements and space utilization. The results demonstrated significant improvements in crowd flow management, with the RWS method showing the highest fitness value despite slower convergence compared to RS. The study highlighted the impact of different methods on the convergence of the multi-objective fitness value based on space elements such as overlapping and standard walkway distances. While both selection methods proved to be effective in optimizing space utilization, the RWS method demonstrated greater computational efficiency while still adhering to standard layout designs. This efficiency helps to ensure smoother evacuation and ease of movement during emergency situations.

Author 1: Najihah Ibrahim

Author 2: Fadratul Hafinaz Hassan

Author 3: Sharifah Mashita Syed-Mohamad

Author 4: Rosmayati Mohemad

Author 5: Ahmad Shukri Mohd Noor

Keywords: Genetic algorithm; optimization; spatial layout arrangement; space utilization; urban planning; facility layout design; rank selection; roulette wheel selection

Paper 57: AutiSim: A Virtual Reality Simulation Game Based on the Autism Spectrum Disorder

Abstract: Technologies with altering reality like virtual reality (VR) have become more relevant to the public for their capabilities in the entertainment and healthcare field, as well as affordable for everyone. However, the emphasis on mental health-related simulation is often ignored due to technical complexities and wrong representation. Therefore, this study leverages the immersive capabilities of VR to create an engaging and educational game experience that simulates the sensory and social challenges faced by individuals with autism spectrum disorder (ASD). The study involves designing and implementing a VR game that places users in various scenarios reflecting the daily experiences of autistic individuals. The VR game aims to educate players about common misconceptions, sensory sensitivities, and social difficulties associated with autism. A literature review on XR technology and ASD was conducted during the pre-production phase to explore past research on autism in video games and shape the overall game vision. The study continues with developing an immersive simulation game using VR with locomotive motion controls and an artificial intelligence non-playable character (AI NPC) with Speech-To-Text function. Finally, the testing phase used two approaches: quantitative analysis, using the System Usability Scale (SUS) to assess usability and the Simulator Sickness Questionnaire (SSQ) to identify discomfort issues such as headaches and blurriness during gameplay, and qualitative analysis, gathering expert’s feedback on the VR game's content and teaching effectiveness.

Author 1: Muhammad Aliff Muhd Farid Arfian

Author 2: Ikmal Faiq Albakri Mustafa Albakri

Author 3: Faaizah Shahbodin

Author 4: Mohd Khalid Mokhtar

Author 5: Asniyani Nur Haidar Abdullah

Author 6: Norhaida Mohd Suaib

Author 7: Muhammad Nur Affendy Nor'a

Author 8: Abdul Hasib Jahidin

Keywords: Virtual reality; simulation game; autism; artificial intelligence

Paper 58: Critical Success Factors for Knowledge Transfer in Enterprise System Projects: A Theoretical and Empirical Investigation

Abstract: Enterprise System Projects (ESPs) are fundamental enablers of digital transformation across organizations, yet they consistently suffer from high failure rates, often attributed to ineffective Knowledge Transfer (KT) practices. Despite the critical role of KT in ensuring project sustainability and long-term organizational learning, limited scholarly attention has been given to identifying and systematically categorizing the success factors that influence KT outcomes in ESPs. The aim of this study is to investigate and conceptualize the Critical Success Factors (CSFs) that influence effective knowledge transfer in ESPs. To address this research gap, a mixed-methods approach is used, combining a literature review with empirical insights from semi-structured interviews with industry practitioners involved in large-scale ESP implementations. The analysis reveals a set of interrelated CSFs that significantly impact KT effectiveness. Some of the key points highlighted are the shared knowledge between cultures, the high expertise of consultants based on technicality and social skills, and the solid and visible management support. These points are integrated into a conceptual framework that enhances conceptual understanding while offering practitioners practical guidance. The study contributes by bridging the gap between the KT concept and ESP implementation, which are connected to the academic discourse, proposing a comprehensive model for successful knowledge transfer during the deployment of ESP. From a practical standpoint, the findings offer organizations a strategic lens to design and implement KT mechanisms that enhance project outcomes and ensure long-term knowledge retention.

Author 1: Jamal M. Hussien

Author 2: Riza bin Sulaiman

Author 3: Ali H Hassan

Author 4: Mansoor Abdulhak

Author 5: Hasan Kahtan

Keywords: Enterprise system projects (ESPs); knowledge transfer (KT); critical success factors (CSFs); digital transformation; knowledge-sharing culture; management support; information systems implementation

Paper 59: Content Validity Assessment Using Aiken’s V: Knowledge Integration Model for Blockchain in Higher Learning Institutions

Abstract: The integration of blockchain technology into higher learning institutions (HLIs) holds the potential to revolutionize data management, enhance transparency, and improve trust in academic systems. However, the effective adoption of blockchain requires a comprehensive and valid model that addresses the specific needs and contexts of HLIs. This study aims to assess the content validity of the Knowledge Integration Model for Blockchain in Higher Learning Institutions using Aiken’s V method. The proposed model was developed through a systematic literature review and refined with expert input. Content validity was evaluated by seven experts with backgrounds in education, blockchain technology, and information systems. Each item within the model was rated for its relevance, clarity, and representativeness using a Likert scale. The instrument, consisting of 50 items across six constructs, was evaluated by seven domain experts using Aiken’s V methodology. Each item was rated based on its relevance and clarity using a 5-point Likert scale. The results revealed that out of 50 items, 21 required revision or removal due to low Aiken’s V scores (<0.70), 21 were deemed acceptable but required minor revisions, and 8 demonstrated strong content validity (V ≥ 0.80). These findings underscore the importance of expert evaluation in refining research instruments and ensuring construct alignment. The use of Aiken’s V provided a robust quantitative foundation for the validation process. The refined instrument serves as a reliable tool for assessing institutional readiness and knowledge integration capabilities in the context of blockchain adoption. This work contributes to the growing research on educational blockchain implementation by offering a validated framework that can support empirical investigations and strategic decision-making in higher education.

Author 1: Nur Ilyana Ismarau Tajuddin

Author 2: Ummu-Hani Abas

Author 3: Khairi Azhar Aziz

Author 4: Rozi Nor Haizan Nor

Author 5: Nor Aziyatul Izni

Author 6: Muhammad Nuruddin Sudin

Author 7: Nur Aqilah Hazirah Mohd Anim

Author 8: Noorashikin Md Noor

Keywords: Blockchain; content validity; aiken’s v; higher learning institutions; knowledge integration model

Paper 60: An Interpretable Transformer-Based Approach for Context-Aware and Stylistically Aligned Academic Paraphrasing

Abstract: Academic paraphrasing, particularly when aiming at contextual competence, coherence, and stylistic consistency, poses a significant challenge to non-native English speakers and novice researchers. This research seeks to create an interpretable transformer model specifically designed for paraphrasing academic texts that guarantees semantic correctness, contextual relevance, and scholarly style. Existing paraphrasing models are largely unsuitable in meeting the subtle needs of academic work, lagging in semantic preservation, fluency, scholarly style, and interpretability. In addressing these limitations, we propose T5-XAVRL (T5 with Attention Visualization and Reinforcement Learning for Style Control), an interpretable Transformer model created specifically for paraphrasing academic text. Based on the T5 architecture, T5-XAVRL adds fine-tuning for better domain adaptation, attention visualization for better transparency, and reinforcement learning to control outputs towards academic writing quality. The model is trained and tested on the ArXiv Academic Papers Dataset and demonstrates high versatility in a variety of academic environments. Developed with Python, TensorFlow, and Hugging Face Transformers, the system is made for scalability as well as performance. Experimental findings indicate that T5-XAVRL obtains a 68.7% BLEU score, greatly surpassing traditional paraphrasing models in both semantic accuracy and linguistic fluency. Far more than a paraphraser, T5-XAVRL is a trustworthy academic writing aide capable of assisting users with producing grammatically and stylistically correct scholarly work. Its interpretable outputs also increase user confidence by vividly displaying how paraphrasing choices are being made. As a whole, this study is an important step towards creating interpretable, context-sensitive, and style-sensitive paraphrasing systems for scholarly use.

Author 1: A. Z. Khan

Author 2: Ritu Sharma

Author 3: K. Kiran Kumar

Author 4: Elangovan Muniyandy

Author 5: Raman Kumar

Author 6: Yousef A. Baker El-Ebiary

Author 7: Prema S

Author 8: Osama R. Shahin

Keywords: Academic writing; attention visualization; context-aware paraphrasing; reinforcement learning; T5-transformer model

Paper 61: Leveraging LSTM-Driven Predictive Analytics for Resource Allocation and Cost Efficiency Optimization in Project Management

Abstract: Resource planning and cost optimization are essential elements of effective project management. Conventional models are weak in changing environments because they cannot keep pace with intricate task interdependencies and changing project constraints. To overcome such weaknesses, this research envisions an LSTM-based predictive analytics model that deploys temporal trends and past project information for precise predictions of task duration, resource allocations, and possible delays. The proposed method combines sequential data modeling with Long Short-Term Memory (LSTM) networks, along with data preprocessing and optimization, to enhance project scheduling and cost control decision-making. With TensorFlow implementation, the proposed LSTM-PRO model resulted in a Mean Squared Error (MSE) of 0.0025, Root Mean Squared Error (RMSE) of 0.05, and an R² score of 0.96, which was far better than ARIMA and other baseline models. The model resulted in a cost saving of 20% on project costs and 20% rise in resource utilization from 65% to 85%. The outcome proves the effectiveness and applicability of the model in actual project settings.

Author 1: G. Gokul Kumari

Author 2: Shokhjakhon Abdufattokhov

Author 3: Sanjit Singh

Author 4: Guru Basava Aradhya S

Author 5: T L Deepika Roy

Author 6: Yousef A.Baker El-Ebiary

Author 7: Elangovan Muniyandy

Author 8: B Kiran Bala

Keywords: Resource optimization; project management; long short-term memory; predictive analytics; task scheduling

Paper 62: AccuLandNet: Enhancing Land Cover Detection with Deep Integrated Learning

Abstract: Now-a-days, population growth is increasing more and more in all the places of the world. Specifically, this increase is in urban development based on economic and industrial improvement. It shows the massive impact on Land Use/Land Cover (LULC) and may change many times. The most popular use of land cover categorization is to analyze satellite imagery to categorize different land surface types, such as urban areas, agricultural fields, forests, and aquatic bodies. With the help of several land cover images, a unique classification model (UCM) based on satellite image classification will be developed in this study. The proposed approach implements the following stages. In the first stage, the pre-trained model U-Net was used to train the satellite images. In the second stage, the preprocessing techniques, including data acquisition and noise reduction, such as Adaptive Noise Removal (ANR) and Histogram Equalization (AHE), were used to preprocess the images. The third stage focused on extracting the features using Multi-Sensor Data Fusion (MSDF) to extract features like water bodies, roads, urban areas, edges, boundaries, and shapes. The final step uses the Maximum Likelihood Classification (MLC) combined with Support Vector Machines (SVM) to give the advanced classification results. Experimental results explain that the proposed approach outperformed the existing models in terms of better outcomes.

Author 1: Geetha Guthikonda

Author 2: M. Senthil Kumaran

Keywords: Land Use/Land Cover (LULC); U-Net; Multi-Sensor Data Fusion (MSDF); Maximum Likelihood Classification (MLC); Support Vector Machines (SVM)

Paper 63: Graph Neural Networks with Attention Mechanisms for Accurate Dengue Severity Prediction

Abstract: Dengue fever continues to be a significant public health issue across the globe because it can lead to life-threatening complications. Severity prediction in a timely and precise manner is imperative for proper clinical management and effective resource utilization. Conventional models fail to identify intricate relationships between heterogeneous clinical, demographic, and epidemiological variables. For this purpose, we develop an innovative framework—Graph Neural Network with Attention Mechanism (GNN-AM)—aimed at enhancing dengue severity prediction. In the suggested method, every patient is viewed as a node in a graph with edges indicating clinical similarity in terms of health properties. The incorporation of attention mechanisms enables the model to selectively pay attention to important clinical indicators like fever duration, platelet count, and bleeding tendencies. This selective attentiveness improves prediction quality by giving maximum importance to the most important features while reducing the impact of less significant data. The model was trained and tested on a dataset of laboratory-confirmed dengue cases that contained clinical symptoms, laboratory results, and demographics. Experimental results showed that attention-augmented GNN performed better than both typical GNNs and traditional machine learning models, recording an accuracy of 90.3%, a recall of 88.9%, and an F1-score of 89.6%. Results highlight the efficacy of the GNN-AM framework in classifying dengue severity accurately and the ability to emphasize crucial clinical indicators using attention mechanisms. In the future, this model can be combined with Electronic Health Records (EHRs) and implemented in real-world healthcare environments using federated learning methods to maintain data privacy across institutions.

Author 1: Monali G. Dhote

Author 2: Puneet Thapar

Author 3: Yousef A. Baker El-Ebiary

Author 4: G. Indra Navaroj

Author 5: R. Aroul Canessane

Author 6: B. V. Suresh Reddy

Author 7: Elangovan Muniyandy

Author 8: Kapil Joshi

Keywords: Attention mechanism; dengue severity prediction Graph Neural Network; healthcare analytics; machine learning

Paper 64: Metaheuristic-Driven Feature Selection for IoT Intrusion Detection: A Hierarchical Arithmetic Optimization Approach

Abstract: The increasing sophistication of cyberattacks in Internet of Things (IoT) networks requires strong Intrusion Detection Systems (IDS) with optimal feature selection mechanisms. High-dimensional data, computational complexity, and suboptimal detection accuracy hinder conventional IDS mechanisms. To overcome these limitations, in this study, the Hierarchical Self-Adaptive Arithmetic Optimization Algorithm (HSAOA) is introduced as a new metaheuristic method for IDS feature selection. HSAOA combines a stochastic spiral exploration method, an adaptive hierarchical model of leaders and followers, and a differential mutation mechanism to improve exploration-exploitation balance, global search capability, and premature convergence. The NF-ToN-IoT dataset is used to test the model, wherein HSAOA undertakes the feature selection process, and classification accuracy is increased by utilizing Random Forest (RF). The experimental results indicate that the proposed HSAOA is better than other advanced approaches in accuracy, computational efficiency, and convergence speed. These results validate the proposed algorithm as a scalable and effective solution for enhancing cybersecurity in IoT environments by improving IDS performance and reducing feature selection complexity.

Author 1: Jing GUO

Author 2: Dejun ZHU

Author 3: Qing XU

Keywords: Intrusion detection; internet of things; feature selection; hierarchical arithmetic optimization; cybersecurity

Paper 65: Enhancing SVM and KNN Performance Through Preprocessing Pipelines for Interactive mHealth Applications

Abstract: Mobile health (mHealth) applications are increasingly relying on artificial intelligence (AI) to provide accurate and real-time decision support for healthcare delivery. However, achieving the optimal balance between processing time and accuracy remains challenging, especially for interactive applications that rely on cloud computing for scalability and performance. This study investigates the impact of data preprocessing techniques on the performance of two widely used machine learning algorithms, Support Vector Machine (SVM) and k-Nearest Neighbors (KNN), in cloud-based mHealth systems. We evaluate the effects of various scaling methods and dimensionality reduction techniques, on processing time and model accuracy. Our results demonstrate that preprocessing significantly improves model performance, with SVM achieving a precision of 0.72 and a processing time of 0.087 ms using StandardScaler, while KNN demonstrates the fastest processing times when paired with robust preprocessing. These findings underscore the importance of optimizing both data preparation and algorithmic efficiency for interactive mHealth applications. By enhancing model accuracy and reducing latency, this research contributes to the development of cost-effective, real-time mobile health systems that improve user experience and decision-making in healthcare.

Author 1: Btissam Elaziz

Author 2: Charaf Eddine AIT ZAOUIAT

Author 3: Mohamed Eddabbah

Author 4: Yassin LAAZIZ

Keywords: Mobile health; cloud computing; machine learning; SVM; KNN; data preprocessing

Paper 66: Deep Reinforcement Learning Based Robotic Arm Control Simulation to Execute Object Reaching Task for Industrial Application

Abstract: This study presents a deep reinforcement learning (DRL) approach to train a robotic arm for object reaching tasks in industrial settings, eliminating the need for traditional task-specific programming. Leveraging the Proximal Policy Optimization (PPO) algorithm for its stability in continuous control, the system learns optimal behaviors through autonomous trial-and-error. Central to this work is reward shaping, where structured feedback based on distance to the target, collision avoidance, motion constraints, and step efficiency guides the agent, akin to incremental coaching. A simulated industrial environment was developed using Webots, integrated with OpenAI Gym and Stable-Baselines3, enabling safe training with sensor data (camera, distance sensor) and randomized target placements. Three models with varying reward schemes were evaluated: simpler rewards prioritized rapid convergence, while complex formulations (e.g., perceptual alignment) enhanced long-term accuracy at the cost of initial instability. Experimental results demonstrated that reward shaping reduced the required steps, highlighting its role in accelerating learning. The study underscores the efficacy of combining DRL, simulation-based training, and adaptive reward design to develop efficient robotic controllers. These findings advance scalable solutions for industrial automation, emphasizing the trade-offs between reward complexity and policy convergence. Future work will refine reward functions to bridge simulation-to-reality gaps, fostering practical adoption in manufacturing and assembly systems.

Author 1: John Mark Correa

Author 2: Rudolph Joshua Candare

Author 3: Junrie B. Matias

Keywords: Reinforcement learning; deep reinforcement learning; reward shaping techniques; robotic arm; robot simulation

Paper 67: Explainable Deep Temporal Modeling for Stroke Risk Assessment Using Attention-Based LSTM Networks

Abstract: Stroke continues to be a major cause of mortality and disability globally, and precise risk prediction models are needed. Current models do not effectively incorporate temporal patient information, restricting the quality of prediction and clinical interpretability. This research introduces a new LSTM-based deep learning model enriched with an attention mechanism for predicting stroke risk that can prioritize important risk factors like age, hypertension, and heart disease. The model takes advantage of LSTM's ability to learn sequential dependencies from long-term patient histories, while the attention mechanism dynamically emphasizes clinically important features, promoting interpretability and clinical significance. By testing the model using a dataset of 5,110 patient records with a mere 6% stroke cases, showcasing extreme class imbalance. To counteract this, preprocessing involved SMOTE for synthetic oversampling, mean imputation to handle missing values, and Min-Max normalization. As deployed in Python based on TensorFlow, the model realized remarkable performance. The constructed LSTM-Attention model attained a test accuracy of 83.7%, an AUC-ROC value of 85.3%, and an F1-value of 82.2%, which was higher than that of conventional models such as Logistic Regression and Random Forest. These evaluate the model's improved ability to identify subtle stroke risk factors that go unnoticed otherwise. The attention-augmented LSTM architecture not only guarantees accurate predictions but also offers transparent insight into the decision process, making it appropriate for incorporation in real-time clinical decision support systems. This method has the potential to improve personalized stroke risk assessment dramatically and enhance preventive healthcare interventions.

Author 1: P. Selvaperumal

Author 2: F. Sheeja Mary

Author 3: Pratik Gite

Author 4: T L Deepika Roy

Author 5: Yousef A. Baker El-Ebiary

Author 6: Gowrisankar Kalakoti

Author 7: Sandeep Kumar Mathariya

Keywords: Attention mechanism; deep learning; imbalanced data; LSTM networks; SMOTE resampling; stroke prediction

Paper 68: Improving Cross-Lingual Fake News Detection in Indonesia with a Hybrid Model by Enhancing the Embedding Process

Abstract: In the digital age, the spread of false information across languages in digital form threatens the authenticity and credibility of information. This study aims to develop an efficient hybrid deep learning model for detecting cross-lingual fake news, particularly in resource-constrained environments, by enhancing the embedding process. It proposes a lightweight model that combines MUSE embeddings with CNN, LSTM, and LSTM-CNN architectures to evaluate performance across various language pairs with Indonesian as the source language. Experiments show that linguistic similarity significantly influences classification performance. CNN achieves an F1-score of 82% for the Indonesian–Malay pair, a similar language pair. While LSTM achieves 97% for the Indonesian–German language pair (a structurally different language pair). These findings highlight the effectiveness of hybrid architectures and multilingual embeddings in improving cross-lingual fake news detection, especially when English is not the source language. The proposed method provides a reliable yet computationally efficient solution for multilingual misinformation detection in resource-constrained environments.

Author 1: Jihan Nabilah Hakim

Author 2: Yuliant Sibaroni

Keywords: Cross-lingual; fake news detection; hybrid learning; MUSE embeddings; digital misinformation

Paper 69: A Preliminary Study on Songket: A Preservation of Intangible Cultural Heritage

Abstract: Songket is a traditional Malaysian woven fabric, opulent, and symbolizes luxurious classical textiles of the old craft in Malaysia. Songket is part of the intangible cultural heritage to this day. However, cultural heritage preservation arises as a critical endeavor, especially for younger generations. There is a growing concern about the disinheritance of cultural heritage to posterity due to disregard for the value of our own heritage. Therefore, this paper aims to explore in-depth insight into the beautiful heritage of songket weaving including the knowledge of art weaving, delve into the techniques and materials used as well as the legacy and time ahead. The research uses observation methods and deep face-to-face interviews for data collection. It was conducted with experienced experts and workers, revealing the origin of Songket and its processing from scratch. The research data findings were gathered and subsequently analyzed as secondary data. This study gives more knowledge about Malay traditional textile art and heritage, indicating a positive future of traditional weaving arts as well.

Author 1: Nik Siti Fatima Nik Mat

Author 2: Syadiah Nor Wan Shamsuddin

Author 3: Syarilla Iryani Ahmad Saany

Author 4: Norkhairani Abdul Rawi

Author 5: Julaily Aida Jusoh

Author 6: Wan Malini Wan Isa

Author 7: Addy Putra Md Zulkifli

Author 8: Shahrul Anuwar Mohamed Yusof

Keywords: Songket; weaving; textiles; aesthetic; intangible cultural heritage; processing songket; cultural; heritage; Terengganu; Malaysia

Paper 70: Utilizing Machine Learning to Identify High-Risk Groups in Sickle Cell Anemia

Abstract: Sickle Cell Anemia (SCA) is a hereditary condition causing abnormal red blood cells, leading to severe health complications. Traditional treatment approaches for SCD often involve reactive management, which can delay appropriate interventions and worsen patient outcomes. The aim of this study is to leverage machine learning (ML) algorithms, including Logistic Regression (LR), Support Vector Machines (SVM), and Decision Trees (DT), to identify high-risk groups among SCA patients using clinical and pathological data from King Abdulaziz University Hospital. This study employs a comprehensive dataset comprising 200 SCA patients, with data preprocessing to handle missing values and feature selection techniques to enhance model performance. The dataset is divided into training and testing sets, and models are evaluated using ten-fold cross-validation. Performance metrics such as True Positive Rate (TPR), False Negative Rate (FNR), Positive Predictive Value (PPV), and False Discovery Rate (FDR) are used to assess model effectiveness. The results indicate that the SVM model with the top seven correlated features achieved the highest TPR and PPV, along with the lowest FNR and FDR, demonstrating its superior performance in identifying high-risk patients. The study concludes that ML models, particularly SVM, can significantly improve risk assessment and patient management in SCA, offering a proactive tool for healthcare providers. The main message is the potential of ML algorithms to enhance clinical decision-making and improve outcomes for patients with SCA.

Author 1: Haneen Banjar

Author 2: Nofe Alganmi

Author 3: Hajar Alharbi

Author 4: Ahmed Barefah

Author 5: Hatem Alahwal

Author 6: Salwa Alnajjar

Author 7: Abdulrahman Alboog

Author 8: Salem Bahashwan

Author 9: Galila Zaher

Keywords: Sickle cells anemia; feature selection; predicting complication; machine learning

Paper 71: Real Time Accident Detection and Emergency Response Using Drones, Machine Learning and LoRa Communication

Abstract: Road accidents and delayed emergency responses remain a major concern in urban environments, contributing to over 1.4 million fatalities globally each year. With rapid urbanization and increasing vehicle density, timely detection and efficient traffic management are critical to reducing the impact of such events. This study proposes a real time Accident Detection and Emergency Response System with integrating Machine Learning IoT enabled drones and LoRa communication. The system combines real time accident detection using CCTV, drone assisted fire detection for post accident scenarios, crime activity monitoring and automated traffic management to reduce congestion and improve public safety. LoRa ensure long range, energy-efficient communication. ML models improve detection accuracy across accidents, fires, crimes and vehicles. Figures and sensor data are analyzed in real time to trigger alerts and assist emergency responders. The system supports scalable integration with existing urban infrastructure, promoting the development of smart city safety frameworks. By minimizing emergency response time, limiting secondary incidents and improving situational awareness, the proposed solution addresses critical gaps in current urban safety systems. It offers a practical, intelligent and adaptive approach to accident mitigation and traffic control in smart cities.

Author 1: Bandara H. M

Author 2: Maduhansa H. K. T. P

Author 3: Jayasinghe S. S

Author 4: Samararathna A. K. S. R

Author 5: Harinda Fernando

Author 6: Shashika Lokuliyana

Keywords: Accident detection; machine learning; IoT; drones; traffic management; LoRa communication

Paper 72: Cybersecurity and the NIST Framework: A Systematic Review of its Implementation and Effectiveness Against Cyber Threats

Abstract: This systematic review evaluates the adoption and effectiveness of the NIST Cybersecurity Framework (CSF) in mitigating cyber threats across diverse sectors. Following PRISMA guidelines, we analyzed studies published between 2015 and 2024 from major academic databases, focusing on the framework's five core functions: Identify, Protect, Detect, Respond, and Recover. Results indicate widespread recognition but uneven adoption—large organizations show strong performance in the Protect and Detect functions, while small and medium-sized enterprises (SMEs) face implementation barriers due to limited resources. The framework's flexibility and risk-based approach are notable strengths, though its voluntary nature and lack of localized standards pose challenges. Compared to ISO/IEC 27001 and COBIT, NIST CSF is more adaptable but less prescriptive. We identify key gaps in empirical validation and sector-specific applications, and recommend future research integrating AI-driven threat detection and regional adaptations.

Author 1: Juan Luis Salas-Riega

Author 2: Yasmina Riega-Virú

Author 3: Mario Ninaquispe-Soto

Author 4: José Miguel Salas-Riega

Keywords: Cyberattacks; small and medium enterprises; risk management; organizational resilience; cyberthreats

Paper 73: Hybrid Detection Framework Using Natural Language Processing (NLP) and Reinforcement Learning (RL) for Cross-Site Scripting (XSS) Attacks

Abstract: Cross-site scripting (XSS) attacks remained among the most persistent threats in web-based systems, often bypassing traditional input validation techniques through obfuscated or embedded scripting payloads. Existing detection models typically relied on static rules or shallow learning techniques, limiting their ability to adapt to evolving attack vectors. This research addressed that gap by developing a hybrid detection framework that integrated natural language processing (NLP) and reinforcement learning (RL) techniques to classify and interpret malicious web inputs. The study aimed to design, develop, and evaluate a system that transformed raw input strings into structured features, trained a deep neural network (DNN) for binary classification, and simulated agent-based learning through policy-driven feedback. The methodology followed the Design Development Research (DDR) framework. Preprocessing involved lowercasing, lemmatization, stopword removal, and TF-IDF vectorization. The trained DNN achieved high accuracy and demonstrated clear boundary separability through PCA and t-SNE visualizations. In the simulation phase, the RL agent optimized its classification policy using cumulative rewards, Q-value heatmaps, and decision contour projections. Results confirmed the system’s capability to generalize across input variations while maintaining interpretability and precision. This framework provided a scalable solution for web application security and demonstrated the effectiveness of semantically guided and policy-aware models for detecting XSS threats.

Author 1: Carlo Jude P. Abuda

Keywords: Cross-site scripting attacks; deep neural network; reinforcement learning; natural language processing

Paper 74: Evaluation Index System for Environmental Restoration Effectiveness Based on Landscape Pattern and Ecological Low-Carbon Construction

Abstract: Traditional green view rate (GVR) methods, which rely on two-dimensional planar images, have several limitations. They fail to capture the three-dimensional spatial characteristics of urban greenery, are frequently dependent on subjective parameters such as camera angles and lighting, and require labor-intensive manual analysis. These factors limit the accuracy and scalability of green space assessments. To overcome these challenges, this study introduces the Panoramic Green Perception Rate (PGPR). This novel metric utilizes spherical panoramic imagery and deep learning for the automated recognition of three-dimensional vegetation. A Dilated ResNet-105 network was used, achieving a mean Intersection over Union (mIoU) of 62.53% with only a 9.17% average deviation from manual annotation. PGPR was empirically applied in Ziyang Park, Wuhan, where it effectively quantified green visibility across urban activity spaces. This approach allows for the scalable and objective evaluation of urban greenery, which has practical applications in urban planning, landscape assessment, and ecological low-carbon construction. Urban planners, environmental engineers, and computer vision and smart city development researchers will find it especially useful.

Author 1: Jingyuan Mao

Keywords: Panoramic green perception rate; deep learning; urban green space; vegetation recognition; landscape assessment

Paper 75: Application Analysis and Research of Text Model Based on Improved CNN-LSTM in the Financial Field

Abstract: With the continuous development of information technology, public opinion analysis based on open-source texts and financial situation awareness has become a research hotspot. This study focuses on financial news and commentary information. First, a topic crawler classification model combining the advantages of CNN and LSTM is proposed to improve the topic recognition ability of financial news texts, and a CNN-LSTM-AM stock price fluctuation prediction model is proposed. This model performs sentiment analysis through BiLSTM, integrates multiple emotional factors and market historical data, and demonstrates superior predictive performance compared to traditional models in multiple experiments.

Author 1: Jing Chen

Author 2: Chensha Li

Keywords: Financial information mining; CNN-LSTM model; stock price prediction; sentiment analysis; BiLSTM

Paper 76: Ensuring Consistency in Group Decision Making: A Systematic Review of the FWZIC Method

Abstract: Subjective opinions in decision-making processes are often vague, ambiguous, and imprecise due to the inherent subjectivity and variability in individual perspectives. This systematic study examines the Fuzzy Weighted Zero Inconsistency (FWZIC) method, which addresses these challenges by achieving consistency in group consensus and effectively managing uncertainties associated with subjective human opinions. The FWZIC method is increasingly popular in the Multi-Criteria Decision Making (MCDM) field for determining criteria weights. This study comprehensively analyzes 71 empirical studies published from 2021 to March 2025, employing the FWZIC method across diverse domains such as healthcare, engineering, and supply chain. By categorizing FWZIC literature based on themes and domains, this study reveals a taxonomy of the latest techniques and methods integrated with FWZIC. It also explores fuzzy extensions and integrated MCDM methods, providing researchers with a summary of suitable techniques for various contexts. By systematically synthesizing findings, this study provides a comprehensive overview of the current state of FWZIC applications in the literature, identifies gaps and suggests potential avenues for future research in the MCDM domain.

Author 1: Ghazala Bilquise

Author 2: Samar Ibrahim

Keywords: FWZIC; MCDM; fuzzy sets; subjective judgment; group decision making

Paper 77: Design and Implementation of Low-Cost Hybrid-Controlled Smart Wheelchair Based on PID Control Integrated with Vital Signs Monitoring

Abstract: According to international organizations’ statistics, the percentage of disabled people is considered not just a small percentage of the world's population. Improving the quality of life for people by using new technologies is one of the essential topics today. Although the wheelchair is the most common way of mobility for the disabled, this research aims to improve wheelchair use by creating other control methods, including speech recognition commands. A joystick, wireless remote control, and an additional port are available in addition to speech recognition commands to control the wheelchair. Researchers also studied and compared the effects of integrating a Proportional–Integral–Derivative (PID) controller with the normal controller during smart wheelchair operation in various normal usage scenarios. This research provides data based on a real experiment, not like most research that depends on mathematical models only for comparison. Adding the PID controller eliminated the overshoot of the smart wheelchair, reduced steady-state error, and reduced settling time. Furthermore, it contains a healthcare monitoring system to track the user's vital signs and object avoidance sensors to keep them safe. Also, a full motor selection calculation for the smart wheelchair has been provided, which is useful for mobile robot design. Additionally, the smart wheelchair features a power monitoring system. Finally, a voice-controlled wheelchair helps the user feel more private and independent. Using this kind of smart wheelchair and living this lifestyle will increase the user's morale.

Author 1: M. Sayed

Author 2: MG Mousa

Author 3: Ali A. S

Author 4: T. Mansour

Keywords: Smart wheelchair; speech recognition; PID; healthcare; mobile robot

Paper 78: Design and Implementation of an Intelligent Laboratory Management System Based on UWB Technology

Abstract: In recent years, the rapid development of educational informatization and the widespread adoption of Internet of Things (IoT) technologies have accelerated the transformation of university laboratories toward intelligent management. However, traditional laboratory management systems still suffer from limited automation, insufficient safety mechanisms, and poor real-time responsiveness. To address these issues, this study proposes an intelligent laboratory management system based on Ultra-Wideband (UWB) technology, which offers high-precision positioning and low-latency communication. The proposed system integrates IoT-based functionalities across six core modules, including access control, asset management, environmental monitoring, and user management. By deploying UWB tags and sensors throughout the laboratory environment, the system enables real-time tracking of personnel and equipment, automatic activation of laboratory devices, and intelligent safety alerts. A pilot deployment in three university laboratories demonstrated improvements in access efficiency, energy conservation, equipment security, and user satisfaction. The results validate the system’s effectiveness in enhancing intelligent laboratory management and provide a scalable model for future smart educational infrastructure.

Author 1: Heng Sun

Author 2: Qiang Gao

Keywords: UWB; intelligent management; laboratory management system; IoT

Paper 79: RFID Integration with Internet of Things: Data Processing Algorithm Based on Convolutional Neural Network

Abstract: Radio Frequency Identification is a fast and reliable communication module that performs automatic data capture to identify and track individual objects and people. Frequency-coded tags employ resonant networks to decode their unique code. A multi-scatterer or multi-resonant method encodes the data. Research primarily related to the current investigation predicted that the chipless RFID tag resonant network has a high bit encoding capacity. This study addresses the simulation, optimization, fabrication, testing, and data encoding methods for chipless RFID tags. This research provides a framework for the open-ended quarter-wavelength stub multi-resonator method in chipless Radio Frequency Identification (RFID) tags. The proposed design enhances the tag's data encoding capacity and improves its robustness to ecological differences. This study integrates Error Correction Coding (ECC) and Adaptive Modulation Systems (AMS) employing Convolutional Neural Networks (CNN) to enhance the tag's performance. The AMS dynamically alters the modulation parameters based on channel states, while ECC improves data reliability. The results indicate efficient performance compared to traditional chipless RFID tags, highlighting the possibility of practical behavior in typical applications that necessitate reliable and high-capacity data transmission.

Author 1: Liang Wang

Keywords: RFID; chipless; coding; threshold; data transmission; error correction; security authentication

Paper 80: A Technique to Support Incremental Construction and Verification in Component-Based Software Development

Abstract: Technological advancements in recent decades have significantly increased the scale and complexity of software systems, which poses challenges to their development and reliability. Component-based software development (CBSD) offers a promising solution by enabling modular and efficient software construction. However, CBSD alone cannot fully address challenges such as ensuring reliability and avoiding errors like deadlocks. Verification techniques, such as model-checking, are necessary to ensure the correctness of CBSD systems. Despite its effectiveness in verifying system properties, model-checking faces a critical issue known as state-space explosion (SSE), which hinders scalability. This study introduces an incremental verification technique for CBSD to address SSE and ensure deadlock freedom. The proposed technique incrementally constructs and verifies component-based systems, eliminating verified portions of components to minimize state-space size during subsequent verification steps. It utilizes a component model that supports encapsulation of computation and control, making incremental verification feasible. Evaluation of the technique using coloured petri nets with non-trivial case studies demonstrates its ability to detect deadlocks early and manage SSE effectively, thereby improving the efficiency of the verification process.

Author 1: Faranak Nejati

Author 2: Ng Keng Yap

Author 3: Abdul Azim Abd Ghani

Keywords: Component-based software development; incremental software construction; software verification

Paper 81: Real-Time Video Captioning on CPU and GPU: A Comparative Study of Classical and Transformer Models

Abstract: This study proposes a scalable and hardware-adaptable approach to automatic video caption generation by comparing two architectures: a traditional encoder–decoder framework combining InceptionResNetV2 with GRU and a transformer-based model integrating TimeSformer with GPT-2. The system supports CPU and GPU deployment through a unified pipeline built on FFmpeg and ImageMagick for keyframe extraction and subtitle embedding. Experimental evaluations on the MSVD and VATEX datasets demonstrate that the TimeSformer–GPT-2 architecture significantly outperforms baseline models, particularly in GPU settings, achieving top results across BLEU, METEOR, ROUGE-L, and CIDEr metrics. This superiority is attributed to its capacity to model spatiotem-poral dependencies and generate contextually rich language. Designed for real-time operation, the system is also suitable for low-resource devices, enabling impactful applications such as assistive tools for the visually impaired and intelligent video indexing. Despite high computational demands and sequence-length limitations, the system presents promising directions for future development, including multilingual captioning, multimodal audio–visual integration, and lightweight models like TinyGPT for enhanced portability.

Author 1: Othmane Sebban

Author 2: Ahmed Azough

Author 3: Mohamed Lamrini

Keywords: Video captioning; transformer; timesformer; GPT-2; real-time inference; spatiotemporal attention; multimedia accessibility; CPU and GPU deployment

Paper 82: Endometriosis Lesion Classification Using Deep Transfer Learning Techniques

Abstract: In resource-limited settings, assisting physicians with disease identification can significantly improve patient outcomes. Early diagnosis is crucial, as many patients could remain healthy with timely intervention. Recent advancements in deep learning models for medical image processing have enabled algorithms to achieve diagnostic accuracy comparable to that of healthcare professionals. This research aims to develop a comprehensive system for the rapid and precise detection of endometriosis lesions. We explore the several deep transfer learning architectures, specifically MobileNetV2, VGG19, and InceptionV3, on the Gynecologic Laparoscopy Endometriosis Dataset (GLENDA). Through extensive literature review and parameter optimization, we find that MobileNetV2 outperforms the other models in terms of accuracy. However, challenges remain, as healthcare imaging datasets often suffer from limited sample sizes and uneven class distributions. Collecting additional samples can be costly and time-consuming, which is a prevalent issue in medical imaging. To address this, we employ Deep convolutional Generative Adversarial Networks (DCGAN) to enhance the dataset by generating synthetic images, thus improving class balance. This image augmentation strategy not only boosts model performance but also reduces the manual effort required for image labeling. We evaluate our proposed model using metrics such as accuracy, precision, recall, and F1-score. Initially, our model achieves an accuracy of 95%. The introduction of synthetic samples results in an increased accuracy of 99%, reflecting a 4%improvement and enhancing the model’s overall efficacy.

Author 1: Shujaat Ali Zaidi

Author 2: Varin Chouvatut

Author 3: Chailert Phongnarisorn

Keywords: Endometriosis classification; lesion detection; medical image classification; deep learning; transfer learning; DCGAN

Paper 83: LASSO-Based Feature Extraction with Adaptive Windowing via DTW for Fault Diagnosis in Rotating Machinery

Abstract: In real-world engineering environments, faults in rotating machines typically occur for concise periods, which leads to poor stability and low accuracy in fault diagnosis. The traditional fault diagnosis of rotating machinery relies on analyzing time-series data to detect system degradation and faulty components. However, the complexity of rotating machinery and the presence of multiple fault types across different operating conditions challenges for conventional classification techniques. This paper proposes a LASSO regression-based feature extraction method with adaptive window based on Dynamic Time Warping (DTW) for fault diagnosis in rotating machinery. The approach effectively extract features by modeling the relationship between shaft rotational speeds (25, 50, and 75 rpm) and vibration signals from piezoelectric accelerometers. This research focus on single and combination faults analysis to include 11 faults, enhancing its applicability to real-world fault conditions. To assess its effectiveness, the proposed method is evaluated against Principal Component Analysis (PCA) and Independent Component Analysis (ICA) using the K-Nearest Neighbors (KNN) classifier. The experimental results demonstrate that the LASSO-based approach consistently achieves high classification accuracy across different speeds, outperforming PCA and ICA in both single and double fault scenarios. These findings highlight LASSO regression as a robust feature extraction technique for improving fault detection and predictive maintenance in rotating machinery.

Author 1: Jirayu Samkunta

Author 2: Patinya Ketthong

Author 3: Nghia Thi Mai

Author 4: Md Abdus Samad Kamal

Author 5: Iwanori Murakami

Author 6: Kou Yamada

Author 7: Nattagit Jiteurtragool

Keywords: Rotating machinery; fault analysis; feature extraction; LASSO regression

Paper 84: The Influence of Familiarity with Traffic Regulations on Road Safety: A Simulated Study on Roundabouts and Intersections

Abstract: International drivers who come from keep-right countries and drive in keep-left countries are frequently involved in road accidents due to unfamiliarity with keep-left traffic regulations. Due to unfamiliarity of the traffic regulation, the driver’s performance and behavior are subject to change. The objective of this study was to explore the effects of familiarity with traffic regulations on driving performance and behavior at roundabouts and intersections. To achieve this in a safe environment, twenty-one male familiar drivers and thirty- four male unfamiliar drivers participated in driving in a simulated keep-left traffic regulation. The factors observed were not fastening the seat belt, entering the driving simulation from wrong side, using an improper approaching lane, not signaling, speeding, driving against the traffic flow and using an improper exiting lane at each roundabout and intersection. Mann-Whitney U test was used to compare driving behavior and performance between the familiarity groups. Unfamiliar drivers made significantly more driving mistakes on roundabouts than unfamiliar drivers. Also, some unfamiliar drivers got inside the vehicle from the passenger side instead of driver side and drove against the traffic flow inside the roundabouts. The implications for familiar and unfamiliar driving can be considered for future research development.

Author 1: Raghda Alqurashi

Author 2: Hasan J. Alyamani

Author 3: Nesreen Alharbi

Author 4: Hasan Sagga

Keywords: Driving behavior; driving performance; familiarity with traffic regulations; road intersections; roundabouts

Paper 85: Predicting Jobs, Shaping Economies: Bibliometric Insights into AI and Big Data in Workforce Demand Analysis

Abstract: The integration of Big Data and Artificial Intelligence (AI) is fundamentally transforming how labor markets are analyzed, predicted and managed. Despite significant advances in using these technologies for workforce analytics, the field suffers from several critical limitations: existing approaches predominantly rely on data from online job portals that may not capture informal employment sectors, current predictive models lack robustness in long-term forecasting under rapid economic transformations and cross-border data integration remains insufficiently addressed for comprehensive global analyses. Moreover, the field lacks a structured, quantitative assessment of scientific production that provides a comprehensive overview of research developments, with most existing studies being case-specific or focusing on narrow applications, leaving significant gaps in understanding the intellectual structure, key contributors and thematic evolution of this interdisciplinary domain. To address these research gaps, this study presents the first comprehensive bibliometric analysis of global scientific research examining the intersection of AI, Big Data and labor market prediction. Drawing on a systematic dataset of 276 publications from Scopus, Web of Science and OpenAlex databases spanning 2003 to 2025, this research employs advanced bibliometric techniques to map the intellectual landscape of this rapidly evolving field. Through a structured four-phase methodological framework incorporating performance analysis, science mapping and thematic evolution, the study identifies research trends, intellectual structures, influential contributors and emerging themes. The analysis reveals significant developments in predictive modeling, natural language processing, and hybrid AI approaches for recruitment forecasting and workforce analytics, while highlighting critical challenges posed by algorithmic bias and ethical considerations in AI-driven systems. Key contributions include: 1) the first systematic scientific mapping of the AI-Big Data-labor market intersection 2) identification of research gaps and future directions for long-term labor market prediction, 3) comprehensive analysis of institutional networks and collaborative patterns and 4) evidence-based recommendations for addressing data integration and model interpretability challenges. The findings offer actionable insights for researchers, policymakers and practitioners seeking to leverage intelligent systems to shape the future of work in the digital economy while addressing current methodological limitations.

Author 1: EL Massi Fouad

Author 2: ELouadi Abdelmajid

Keywords: Big data; Artificial Intelligence; predictive modeling; bibliometric analysis; natural language processing; labor market analytics

Paper 86: Dynamic Polygon-Based Reverse Driving Detection Technique for Enhanced Road Safety

Abstract: Reverse driving and lane collapse pose serious risks to road safety, especially on complex infrastructures such as multi-lane highways, intersections, and roundabouts. Existing detection systems often depend on rigid lane configurations and struggle to adapt to varied road geometries and environmental conditions. Prior works are typically limited to straight, multi-lane roads and rely on automated boundary extraction, making them unsuitable for irregular traffic layouts. To address this gap, the objectives of this research paper is to propose a vision-based detection system that combines the YOLOv8 object detector with a dynamic polygon-based zone management strategy. The system aims to detect reverse driving and lane collapse incidents in real time using CCTV footage, without requiring additional sensors. Its key novelty lies in manually configurable zones and the integration of ByteTrack for robust vehicle tracking across complex scenes. The system was tested under diverse real-world parameters, including different road types (single-lane, multi-lane, roundabouts), lighting conditions (day and night), and traffic behaviors (normal flow, reverse, and collapse) and visual evaluations highlight consistent and logically coherent results across scenarios, highlighting its practical effectiveness for real-time intelligent traffic monitoring.

Author 1: Tara Kit

Author 2: Youngsun Han

Author 3: Anand Nayyar

Author 4: Tae-Kyung Kim

Keywords: Reverse driving detection; lane collapse detection; polygon zones; object detection; YOLOv8

Paper 87: Comparative Analysis of Machine Learning Frameworks for Robust Ovarian Cancer Detection Using Feature Selection and Data Balancing

Abstract: One of the most serious malignancies that affects women’s health worldwide is ovarian cancer. As a result, prompt accurate diagnosis and treatment are necessary. This study’s primary objective is to determine whether or not OC is present within the body of a person by using a range of characteristics gleaned with a couple of health examinations. The article is concentrated on twelve ML techniques used for OC diagnosis. The dataset has been altered by applying the borderline SVMSMOTE method to address the imbalance properties and the MICE imputation method to impute the missing values in order to enhance the performance of the classifiers. Addition-ally, the boruta approach and recursive feature reduction has been utilized to identify the most important features while the hyper parameter tuning strategy has been employed to improve classifier performance and provide ideal solutions.Boruta opted just 50% of the total characteristics and outperformed RFE while considering the most important feature. Furthermore, many performance measures are used to determine which classifiers are the best in identifying OC. Voting classifier surpassed state-of-the-art approaches and other machine learning methods with the highest accuracy. The suggested approach obtained the highest average of 93.06% accuracy, 88.57% precision, 96.88% recall, 92.54% F1-score, and 93.44% AUC-ROC based on experimental results. Experiments show that in comparison with the state-of-the-art techniques, our suggested method can identify OC more accurately.

Author 1: DSS LakshmiKumari P

Author 2: Maragathavalli P

Keywords: Ovarian cancer detection; machine learning frame-work; data balancing; feature selection

Paper 88: Comparative Study of Prenatal and Postnatal Images for Detecting Down Syndrome in Children

Abstract: Down syndrome is a genetic disorder caused by the presence of an extra copy of chromosome 21, affecting both neurological development and physical features. Early and accurate diagnosis is critical for ensuring timely medical intervention and support. This study presents a comparative analysis of prenatal (ultrasound) and postnatal (facial) imaging modalities for the detection of Down syndrome using deep learning techniques. We employed VGG19, ResNet50, DenseNet121, MobileNetV2, and the Vision Transformer for image classification. An ensemble model integrating four CNN architectures achieved superior performance, with 92% test accuracy on prenatal images and 83%on postnatal images. Among the individual models, ResNet50 out-performed the others across both modalities. Evaluation metrics, including accuracy, precision, recall, and F1-score, confirm the effectiveness of the proposed framework. These results highlight the potential of ensemble learning to enhance the early detection of Down syndrome and improve accessibility to healthcare.

Author 1: Labanti Singha

Author 2: Iqbal Ahmed

Keywords: Down syndrome; prenatal ultrasound; postnatal facial recognition; CNN; vision transformer; ensemble learning

Paper 89: Sign3DNet: An Enhanced 3D CNN Architecture for Bengali Word-Level Sign Language Recognition

Abstract: Automated recognition of sign languages has been playing an important role in breaking barriers to communication and inclusion for the deaf and mute community. Several studies have been conducted on Bengali Sign Language (BdSL). However, Bengali Word-Level Sign Language (BdWLSL) remains unexplored due to the lack of large annotated datasets and a stable model. Therefore, in this research, we introduced a large-scale Bengali word-level video dataset and proposed a modified 3D Convolutional Neural Network (CNN) architecture for word-level BdSL recognition, emphasizing its ability to capture the spatial and temporal dynamics from video data. The proposed strategy represents strong performance in Bengali word-level sign language recognition by utilizing the spatiotemporal pattern captured by the modified 3D CNN architecture. The proposed model demonstrates its potential for practical use by successfully learning complex hand movements straight from raw video data. The proposed CNN model is benchmarked against traditional deep learning techniques, Temporal Shift Module (TSM), Long Short-Term Memory (LSTM), and default 3D-CNN, providing a comprehensive comparison of their strengths and limitations. Experiments are conducted using a structured video dataset containing 102 Bengali sign-word classes. To ensure privacy, the volunteers’ faces were blurred and only landmark data extracted using MediaPipe, rendered on black backgrounds, were used for training. The experimental result analysis shows that the performance of the proposed 3D-CNN model achieves a satisfactory accuracy of 58.25%, demonstrating its potential for word-level sign language recognition tasks. To our knowledge, this is the very first pilot study for BdWLSL recognition. Hence, we consider the recognition rate 58.25% of the proposed modified 3D-CNN architecture to be satisfactory and a potential scope for future researchers in the same field.

Author 1: Safi Ullah Chowdhury

Author 2: Nasima Begum

Author 3: Tanjina Helaly

Author 4: Rashik Rahman

Keywords: Bengali sign word recognition; computer vision; deep learning; convolutional neural network; spatial-temporal dynamics; video data

Paper 90: Integrating Blockchain and Smart Card Technologies for Secure Healthcare Data Management

Abstract: In recent years, the healthcare sector has faced growing challenges in managing patient data securely and efficiently, especially when it comes to data privacy and the way information is shared across healthcare providers. A number of digital solutions have been proposed over time, but more recently, blockchain has started to gain serious interest. Its structure allows data to remain intact and traceable, while also offering a strong layer of security. This paper explores how blockchain based smart contract might be used alongside smart cards to offer a more robust system for protecting patient information. Smart cards bring in a physical barrier that helps limit access to only those who are authorized, while blockchain makes it much harder to tamper with information or centralize control. The suggested method demonstrates how the decentralized and immutable nature of blockchain, combined with the physical authentication provided by smart cards and the automation of smart contracts improve data security and restrict unauthorized access. The proposed framework is evaluated through smart contract deployment and testing on both the Hardhat local network and the Celo public testnet. The results confirm the practicality and efficiency of the solution and support its potential for real world application in secure healthcare data management.

Author 1: Zayneb Gaouzi

Author 2: Imad Bourian

Author 3: Khalid Chougdali

Keywords: Healthcare; security; blockchain; smart contracts

Paper 91: Enhanced Feature Extraction for Accurate Human Action Recognition

Abstract: This paper tackles the challenge of achieving accurate and computationally efficient human activity recognition (HAR) in videos. Existing methods often fail to effectively balance spatial details (e.g. body poses) with long-term temporal dynamics (e.g. motion patterns), particularly in real-world scenarios characterized by cluttered backgrounds and viewpoint variations. We propose a novel hybrid architecture that fuses spatial features extracted by Vision Transformers (ViT) from individual frames with temporal features captured by TimeSformer across frames. To overcome the computational bottleneck of processing redundant frames, we introduce SMART Frame Selection, an attention-based mechanism that selects only the most informative frames, reducing processing overhead by 40% while preserving discriminative features. Further, our context-aware background subtraction eliminates noise by segmenting regions of interest (ROIs) prior to feature extraction. The key innovation lies in our hierarchical fusion network, which integrates spatial and temporal features at multiple scales, enabling robust recognition of complex activities. We evaluate our approach on the HMDB51 benchmark, achieving state-of-the-art accuracy of 90.08%, out-performing competing methods like CNN-LSTM (85.2%), GeoDe-former (88.3%), and k-ViViT (89.1%) in precision, recall, and F1-score. Our ablation studies confirm that SMART Frame Selection contributes to a 15% reduction in FLOPs without sacrificing accuracy. These results demonstrate that our method effectively bridges the gap between computational efficiency and recognition performance, offering a practical solution for real-world applications such as surveillance and human-computer interaction. Future work will extend this framework to multi-modal inputs (e.g. depth sensors) for enhanced robustness.

Author 1: Tarek Elgaml

Author 2: Ali Saudi

Author 3: Mohamed Taha

Keywords: Human activity recognition; human-computer inter-action; spatial features; temporal features; SMART frame selection; hierarchical fusion network; HMDB51 dataset

Paper 92: Deep Learning in Cephalometric Analysis: A Scoping Review of Automated Landmark Detection

Abstract: Cephalometric landmark identification is funda-mental for accurate cephalometric analysis, serving as a corner-stone in orthodontic diagnosis and treatment planning. However, manual tracing is a labor-intensive process prone to inter-observer variability and human error, highlighting the need for automated methods to improve precision and efficiency. Recent advances in Deep Learning have enabled automatic detection of cephalometric landmarks, thereby increasing accuracy and consistency while reducing processing time. This scoping review examines contemporary applications of Deep Learning in cephalometric landmark detection and cephalometric analysis from 2019 to January 2025. We searched IEEEXplore, Sci-enceDirect, arXiv, Springer, and PubMed databases, identifying 601 articles, of which 76 met inclusion criteria after rigorous screening. Our analysis revealed significant performance improvements with Deep Learning methods achieving Success Detection Rates (SDR) of 75-90% at 2mm thresholds, substantially out-performing traditional methods. Geographical analysis identified China, South Korea, and the United States as leading research centers, with commercial applications like WebCeph and CephX gaining clinical adoption. Deep Learning improves the accuracy and efficiency of cephalometric analysis; however, challenges persist regarding dataset standardization and clinical validation. These technologies show promising potential to support novice clinicians, streamline radiological examinations, and improve landmark identification reliability in routine orthodontic practice.

Author 1: Idriss Tafala

Author 2: Fatima-Ezzahraa Ben-Bouazza

Author 3: Aymane Edder

Author 4: Oumaima Manchadi

Author 5: Bassma Jioudi

Keywords: Artificial Intelligence; deep learning; cephalometric analysis; landmark detection

Paper 93: Advancing Traffic Sign Detection with Convolutional Neural Networks: A Deep Learning Approach

Abstract: Traffic sign detection is a key task in intelligent transportation systems, supporting road safety and traffic flow. This study introduces RoadNet, a lightweight Convolutional Neural Network (CNN) designed for real-time detection and classification of traffic signs in Moroccan road environments. The system addresses challenges such as occlusion, illumination variability, and diverse sign structures. Built on deep learning techniques, RoadNet leverages multiscale feature extraction and transfer learning to improve detection accuracy and generaliza-tion. The dataset includes four sign categories: speed limit, stop, crosswalk, and traffic light. Extensive image preprocessing and augmentation were applied to increase robustness. Results show that RoadNet outperforms baseline models like VGG16, achieving 96% training accuracy and 88.6% validation accuracy, with superior precision, recall, and F1-score. The model maintains low loss and performs reliably under constrained resources. This research confirms the effectiveness of CNN-based architectures for traffic sign detection in real-world Moroccan settings. It contributes to the deployment of AI-powered solutions for smart mobility and logistics, especially in regions with limited computational resources.

Author 1: OUAHBI Younesse

Author 2: ZITI Soumia

Keywords: Traffic sign detection; convolutional neural net-works; deep learning; road safety; intelligent transportation systems; real-time detection; artificial intelligence; transportation efficiency

Paper 94: MITG-CU: Multimodal Interaction Temporal Graphs Approach for Conversational Emotion Recognition

Abstract: In the emotion recognition of conversations, the complementary relationship between the context information and multimodal data cannot be fully exploited. This results in insufficient comprehensiveness and accuracy in emotion recognition. To address these challenges, this paper proposed a Multimodal Interactive Temporal Graph Conversation Understanding model (MITG-CU) composed of textual, audio and visual modalities. Firstly, the pre-extracted textual, audio, and visual features are adopted as the input of the Transformer, and the attention mechanism is utilized to capture the cross-modal context correlation information. Furthermore, structural relationships and temporal dependencies between utterances are captured through a local-level relational temporal graph module. Inter-modal interaction weights are dynamically adjusted by a global-level pairwise cross-modal interaction mechanism. By integrating two complementary hierarchical structures, a hierarchical multimodal information fusion was achieved, and at the same time, the model’s adapt-ability to complex conversation scenarios was enhanced. Finally, feature fusion is carried out by using the gating mechanism and sentiment classification is conducted. Experimental results demonstrated that the proposed model outperforms six common baseline methods across metrics including accuracy, precision, recall, and F1-score. Especially in Weighted-F1 and Accuracy have improved by 0.28 % and 0.39 % respectively, which confirmed the effectiveness of the model.

Author 1: Qian Xing

Author 2: Yaqin Qiu

Author 3: Minglu Chi

Author 4: Xuewei Li

Author 5: Changyi Gao

Keywords: Emotion recognition; multimodal interaction; relational temporal graph; cross-modal interaction; feature fusion

Paper 95: Cross-Domain Evaluation of Large Language Models for Abstractive Text Summarization: An Empirical Perspective

Abstract: Large Language Models (LLMs) have demon-strated remarkable capabilities in generating human-like text; however, their effectiveness in abstractive summarization across diverse domains remains underexplored. This study conducts a comprehensive evaluation of six open source LLMs across four datasets: CNN / Daily Mail and NewsRoom (news), SAMSum (dialogue) and ArXiv (scientific) using zero shot and in-context learning techniques. Performance was assessed using ROUGE and BERTScore metrics, and inference time was measured to examine the trade-off between accuracy and efficiency. For long documents, a sentence-based chunking strategy is introduced to overcome context limitations. Results reveal that in-context learning consistently enhances summarization quality, and chunking improves performance on long scientific texts. The model performance varies according to architecture, scale, prompt design, and dataset characteristics. The qualitative analysis further demonstrates that the top-performing models produce summaries that are coherent, informative, and contextually aligned with human-written references, despite occasional lexical divergence or factual omissions. These findings provide practical insights into designing instruction-based summarization systems using open-source LLMs.

Author 1: Walid Mohamed Aly

Author 2: Taysir Hassan A. Soliman

Author 3: Amr Mohamed AbdelAziz

Keywords: Large language models; natural language processing; automatic text summarization; prompt engineering; summarization evaluation

Paper 96: Foreign Key Constraints to Maintain Referential Integrity in Distributed Database in Microservices Architecture

Abstract: In the world of modern software development, microservices architecture has become increasingly popular due to its ability to help developers to build large and complex applications that are more agile, faster and more scalable. In large scale applications (such as e-commerce, healthcare, finance, social media, inventory management, travel booking, content management, and customer relationship management systems etc.) with many interconnected services, it is tough to keep the data accurate and consistent. The concept of referential integrity is applied to validate the data. Referential integrity refers to the preservation of relationships between tables. In a monolithic architecture, where the application and database are closely linked and co-located on the same server, referential integrity via foreign key constraints makes it feasible to preserve consistent and accurate data. But in the microservices architecture, maintaining referential integrity across distributed databases poses significant challenges due to its decentralized nature of data management. This study utilizes a hybrid research methodology, combining empirical research and design science research to discover and address the challenges of maintaining referential integrity in distributed databases in microservice architecture and calculate the response time by comparing and analyzing with existing models. The results of the evolution in term of response time are presented in this work.

Author 1: Shamsa Kanwal

Author 2: Nauman Riaz Chaudhry

Author 3: Reema Choudhary

Author 4: Younus Ahamad Shaik

Author 5: Pankaj Yadav

Author 6: Ayesha Rashid

Keywords: Foreign key constraints; relational mapping; referential integrity; saga pattern; event driven architecture; APIs; microservice; distributed database

Paper 97: Integrating cGAN-Enhanced Prediction with Hybrid Intervention Recommendations Systems for Student Dropout Prevention

Abstract: Early-warning dashboards in higher education typically stop at tagging students as “at-risk,” offering no concrete guidance for remedial action; this limitation contributes to the loss of thousands of learners each year. Approach. We propose an integrated framework that (i) uses a class-balanced Conditional GAN to augment sparse attrition data, and (ii) couples the resulting XGBoost predictor with a four-mode intervention engine—rule-based, few-shot, fine-tuned LLM, and a novel hybrid strategy—to recommend personalised support. Major findings. Training on GAN-augmented records raises prediction accuracy to 92.79% (a 15.46-point gain over non-augmented baselines), while the hybrid intervention generator attains 94% categorical coverage and the highest specificity score (0.63) albeit at a per-student latency of 61s. Impact. By uniting robust risk prediction with high-quality, actionable interventions, the framework closes the long-standing gap between detection and response, furnishing institutions with a scalable path to materially reduce dropout rates across diverse educational settings.

Author 1: Hassan Silkhi

Author 2: Brahim Bakkas

Author 3: Khalid Housni

Keywords: Student dropout prediction; machine learning in education; personalized intervention systems; Conditional Generative Adversarial Networks(cGAN); Large Language Models (LLMs); hybrid recommendation systems

Paper 98: Habitat Intelligence: How Machine Learning Reveals Species Preferences for Ecological Planning and Conservation

Abstract: The emerging confluence between artificial intelligence and ecology has generated a new research frontier, which we refer to as habitat intelligence, aiming to unveil species environment relationships through data-driven approaches. This SLR aims to summarise the pass to the current year (2025) of the research on the use of ML and DL models to represent species preferences, habitat suitability and ecological niches. Based on 365 peer-reviewed studies extracted from SCOPUS, Web of Science and OpenAlex, we identify four main areas of innovation which encompass: automated species identification and ecological monitoring; AI-enhanced species distribution models (SDMs); advanced data collection and processing for ecological research; and conservation-oriented decision support systems. Our review shows that AI has the potential for a more precise and scalable approach to biodiversity investigations in the age of integrated remote sensing, acoustics, citizen science, and environmental data. But we also point out pressing challenges such as data paucity, model interpretability and computational limitations. We suggest that future advancements in this branch of the food web could come from interdisciplinary cooperation using explainable AI (xAI) and the construction of bridging hybrid models between prediction and ecological interpretability. In the end, this review offers a conceptual and methodological ‘roadmap’ to other researchers and conservation practitioners who wish to apply AI to the service of global biodiversity aims.

Author 1: Meryem Ennakri

Author 2: Soumia Ziti

Author 3: Mohamed Dakki

Keywords: Artificial Intelligence; machine learning; deep learning; species preferences; habitat suitability modeling; species distribution models (SDMs); ecological niche modeling; conservation planning; environmental monitoring; explainable AI (xAI); habitat inte

Paper 99: Deep Learning-Based Bone Age Growth Disease Detection (BAGDD) Using RSNA Radiographs

Abstract: Radiological bone age assessment is essential for diagnosing pediatric growth and developmental disorders. The conventional Greulich-Pyle Atlas, though widely used, is manual, time-intensive, and prone to inter-observer variability. While deep learning methods such as Convolutional Neural Networks (CNNs) offer automation potential, most existing models rely on transfer learning from natural image datasets and lack specialization for medical radiographs. This study aims to address the gap by developing a domain-specific, custom CNN for pediatric bone age prediction. This research proposes a customized CNN architecture trained on the RSNA pediatric bone age dataset, which includes over 12,000 annotated hand X-ray images labeled with age and gender. The pipeline incorporates pre-processing techniques such as image resizing, normalization, and Contrast Limited Adaptive Histogram Equalization (CLAHE) to enhance input quality. A YOLOv3 object detector is utilized to localize the hand region prior to model training, focusing on the most relevant anatomical structures. Unlike traditional transfer learning models such as ResNet50, VGG19, and InceptionV3, the proposed CNN is tailored for radiographic features using optimized convolutional blocks and domain-aware augmentations. This design improves generalization and reduces overfitting on small or imbalanced subsets. The proposed model achieved a Mean Absolute Error (MAE) of 3.27 months on the test set and 3.08 months on the validation set, outperforming state-of-the-art transfer learning approaches. These results demonstrate the model’s potential for accurate and consistent bone age estimation and highlight its suitability for integration into clinical decision-support systems in pediatric radiology.

Author 1: Muhammad Ali

Author 2: Muhammad Faheem Mushtaq

Author 3: Saima Noreen Khosa

Author 4: Naila Kiran

Author 5: Humaira Arshad

Author 6: Urooj Akram

Keywords: Bone age estimation; pediatric healthcare; convolutional neural networks; transfer learning; YOLOv3; medical imaging

Paper 100: Intelligent Agents in Disaster Risk Management: A Systematic Review of Advances and Challenges

Abstract: Artificial Intelligence (AI) has emerged as a trans-formative technology in the domain of Disaster Risk Management (DRM), offering new possibilities for forecasting, preparedness, and rapid response in the face of increasingly frequent and complex natural disasters. This systematic literature review synthesizes the state-of-the-art advances in AI-driven intelligent agents applied to DRM, covering domains such as early warning systems, geospatial analysis, damage assessment, evacuation planning, and decision support. It critically examines the technological innovations, implementation methods, and interdisciplinary approaches that have shaped the evolution of intelligent agent-based solutions in disaster scenarios. Through the analysis of over 7,800 scientific publications indexed in Scopus, Web of Science, and OpenAlex between 2010 and 2025, the review identifies key patterns, application domains, and persistent gaps such as data scarcity, lack of model interpretability, and limited operational deployment. The study also addresses ethical concerns related to AI deployment in high-stakes environments and proposes a roadmap for future integration of intelligent agents with IoT, UAVs, and real-time decision infrastructures. The findings contribute to a deeper understanding of how AI and multi-agent systems can reinforce disaster resilience and inform sustainable and adaptive disaster management strategies at both global and local levels.

Author 1: Hssaine Hamid

Author 2: ELouadi Abedlmajid

Keywords: Intelligent agents; artificial intelligence; disaster risk management; predictive analytics; resilience; early warning systems; geospatial AI; disaster response; ethical challenges; ma-chine learning; climate change adaptation

Paper 101: Artificial Intelligence in Disaster Risk Management: A Scientometric Mapping of Evolution, Collaboration, and Emerging Trends (2003–2025)

Abstract: Recent years have seen a dramatic increase in the number of and severity of natural disasters, driven in part by climate change and urbanization. Artificial Intelligence (AI) appears to be a promising new technology that can transform disaster risk management (DRM) and provide new opportunities for prediction, monitoring, response, and recovery. The present study performs a bibliometric review of applications of AI to DRM, from a total collection of 7842 scientific articles extracted from Scopus, Web of Science and OpenAlex databases from the year 2003 to the year 2025. Exploring the trends of publications, authorship, international collaboration, and research topics, the study reveals the development and current status of AI incorporating disaster management. The results illustrate an apparent growth in interest in the field of science, how machine learning and deep learning methodologies are leading, and the raise of geospatial AI, remote sensing, and social media analysis in disaster preparedness and response. Other issues including data quality, ethics, technology and trust in AI systems are also considered. This study offers helpful perspectives on the status quo and future development of AI-based DRMs.

Author 1: Hssaine Hamid

Author 2: ELouadi Abedlmajid

Keywords: Artificial Intelligence; disaster risk management; machine learning; deep learning; remote sensing; bibliometric analysis; natural disasters; geospatial AI; early warning systems

Paper 102: XPathia: A Deep Learning Approach for Translating Natural Language into XPath Queries for Non-Technical Users

Abstract: XPath is a widely used language for navigating and extracting data from XML documents due to its simple syntax and powerful querying capabilities. However, non-technical users often struggle to retrieve the needed information from XML files, as they lack knowledge of XML structures and query languages like XPath. To address this challenge, we propose XPathia, a novel deep learning-based model that automatically translates natural language questions into corresponding XPath queries. Our approach employs supervised learning on an annotated XML dataset to learn accurate mappings between natural language and structured XPath expressions. We evaluate XPathia using two standard metrics: Component Matching (CM) and Exact Matching (EM). Experimental results demonstrate that XPathia achieves a state-of-the-art performance with an accuracy of 25.85% on the test set.

Author 1: Karam Ahkouk

Author 2: Mustapha Machkour

Keywords: Deep learning; XML databases; neural networks; text-to-XPATH; natural language processing

Paper 103: Random Forest Model Based on Machine Learning for Early Detection of Diabetes

Abstract: Diabetes mellitus presents a growing prevalence at the global level, representing a significant public health challenge. Despite the availability of specific treatments, it is imperative to develop innovative strategies that optimize early detection and management of the disease. The research aims to develop a model that allows for the early detection of diabetes using the Random Forest algorithm, using the Knowledge Discovery in Databases (KDD) methodology, which comprises the phases of selection, preprocessing, transformation, data mining, interpretation and evaluation. The dataset used include 520 randomly selected patient records. The model achieved robust performance, with an accuracy of 85%, sensitivity of 75%, and an F1-score of 78%, indicating an adequate balance between precision and sensitivity. Specificity was 78%, while the area under the ROC curve (AUC) reached 86%, demonstrating a high discriminative ability between positive and negative cases. The balanced accuracy was 82%, and the Matthews correlation coefficient (MCC) registered a value of 0.72, confirming the strength and reliability of the model even in the presence of class imbalance. These results demonstrate the effectiveness of the machine learning-based approach for the early detection of diabetes mellitus, with potential application in clinical decision support systems.

Author 1: Inooc Rubio Paucar

Author 2: Cesar Yactayo-Arias

Author 3: Laberiano Andrade-Arenas

Keywords: Data mining; decision tree; diabetes mellitus; machine learning; random forest

Paper 104: Phishing Simulation as a Proactive Defense: A Customizable Platform for Training and Behavioral Analysis

Abstract: Phishing is one of the most persistent threats, but a lot of awareness programs still use generic, static training. This paper fills in the gap identified above by existing studies through the introduction of a phishing simulation platform that provides personalized, role-based simulation with real-time behavioral tracking. It is a multi-channel delivery (Email, Short Message Service (SMS), (WhatsApp) and can dynamically generate messages using placeholders to simulate realistic attack scenarios. User interactions are visualized on an integrated dashboard to let organizations judge the individual’s risk and provide immediate awareness feedback. Due to ethical restrictions, real user testing could not be performed, and the system was tested using simulated data found to work with a cloud-ready front end. The solution shows great potential for being adopted by enterprises due to its potential to adopt an approach towards cybersecurity training in a more adaptive and engaging way.

Author 1: Abdulrahman Alsaqer

Author 2: Hussain Almajed

Author 3: Khalid Alarfaj

Author 4: Mounir Frikha

Keywords: Phishing; simulation; awareness; analytics; cyber-security

Paper 105: Reducing Computational Complexity in CNNs: A Focus on VGG19 Pruning and Quantization

Abstract: The Convolutional Neural Network (CNN) models are effective in computer vision strategies and have gained popularity due to their strong performance in visual tasks. Nevertheless, models with architectures such as VGG19 are expensive in terms of computational resources and require huge memory, which limits their usage on low-end devices. The study examines how efficiency can be increased in the model VGG19 by using model compression techniques, like pruning (structured and un-structured) and quantization (8-bit and 4-bit Quantization-Aware Training - QAT). The efficiency of the individual compression approaches was tested by thoroughly exploring the VGG19 with the MNIST, CIFAR-10, and Oxford-IIIT Pet datasets. Each model was evaluated against the baseline based on measures of accuracy, model size, inference time, and complexities of the model, CPU usage, and memory usage. The applied QAT approach reduced the model size by 75% with a drop in computational cost across all methods. In addition, the 8-bit quantitative assessment allowed for substantial system compression alongside increased speed delivery with minimal impact on accuracy. The highest compression and sparsity achieved by 4-bit QAT was 48%, which was not effective as it reduced accuracy on complex datasets, with additional computational overhead on T4 GPU. Structured pruning resulted in faster inference, but unstructured pruning also demonstrated a good result in retaining accuracy and even improving it. To simplify the VGG19 structure, pruning and quantization mechanisms are suggested in order to simplify the architecture to implement the model on edge devices sufficiently, without compromising prediction performance.

Author 1: Md. Mijanur Rahman

Author 2: Anik Datta

Author 3: Md. Sabiruzzaman

Author 4: Md Samim Ahmed Bin Hossain

Keywords: VGG19 Model optimization; model compression; pruning; quantization; structured pruning; unstructured pruning; memory management; quantization-aware training; 8-bit; 4-bit

Paper 106: Fake News Detection on Kashmir Issue Using Machine Learning Techniques

Abstract: Focusing events are sudden, impactful occurrences that spark widespread discussions. Analyzing fake news during such events is challenging due to limited and short-lived datasets. Online fact checkers are slow in identifying fake news, and internet communities and forums become the primary source of news, allowing unchecked dissemination. This study proposes a machine learning approach to predict fake news during the revocation of Article 370 in Kashmir as a focusing event. Small dataset from 20th August till 2nd September is collected and user profile parameters are utilized for effective classification. Five classifiers were employed, with Random-Forest and Logistic-Regression achieving the highest F1 scores of 74 per cent. Results identifies prevalent words in true and false news tweets, aiding in fake news detection. This approach mitigates misinformation during events with limited data, contributing to a reliable online environment. The research is valuable for major geopolitical shifts, natural disasters, and social movements.

Author 1: Misbah Kazmi

Author 2: Sadia Nauman

Author 3: Sadaf Abdul Rauf

Author 4: S. Ali

Author 5: Ali Daud

Author 6: Bader Alshemaimri

Keywords: Classification algorithm; fake news; Kashmir issue; machine learning techniques

Paper 107: Text Classification Using Enhanced Binary Wind Driven Optimization Algorithm

Abstract: Document classification using supervised machine learning is now widely used on the internet and in digital libraries. Several studies have focused on English-language document classification. However, Arabic text includes high variation in its morphology, which leads to high extracted features and increases the dimensionality of the classification task. Towards reducing the curse of dimension in Arabic text classification, a wrapper feature selection method is proposed in this study. In more detail, a hybrid metaheuristic model based on the Wind Driven and Simulated Annealing is designed to solve FS task in Arabic text, known as WDFS. The Wind Driven method is initially introduced to optimize the Fs task in the exploration phase. Then, WD is hybridized with simulated annealing as a local search in the exploitation phase to enhance the solutions located by the WD. Three classifiers are utilized to evaluate the selected features using the proposed WDFS: K-nearest Neighbor, Naïve Bayesian, and Decision Tree. The proposed WDFS method was assessed on selected four groups of files from a benchmark TREC Arabic text newswire dataset. Comparative results showed that the WDFS method outperforms other existing Arabic text classification methods in term of the accuracy. The obtained results reveal the high potentiality of WDFS in reliably searching the feature space to obtain the optimal combination of features.

Author 1: Jaffar Atwan

Author 2: Mohammad Wedyan

Author 3: Ahmad Hamadeen

Author 4: Qusay Bsoul

Author 5: Ayat Alrosan

Author 6: Ryan Alturki

Keywords: Text classification; Arabic documents; wind driven optimization algorithm; simulating annealing; feature selection

Paper 108: Comparison of Conventional Techniques for House Electricity Consumption Forecasting

Abstract: Electricity consumption monitoring is the auto-mated process of recording, processing, and analyzing electricity usage in real time to make informed decisions. This research aims to implement an artificial intelligence- and deep learning-based methodology to forecast monthly electricity consumption in Tacna, Peru, and generate decision-making indicators. To this end, we used electricity consumption records from Electrosur S.A., the company responsible for electricity distribution and marketing in the departments of Tacna and Moquegua, from February 2015 to December 2022 (a total of 95 months). We compared three artificial intelligence models in this context: i) eXtreme Gradient Boosting (XGBoost), ii) Light Gradient Boosting (LGBM), and iii) Prophet. While all models effectively fore-casted electricity consumption, the Prophet model demonstrated superior performance, achieving a mean absolute percentage error (MAPE) of 0.7% compared to actual consumption values. Additionally, the study discusses the potential of recurrent neural networks to further enhance predictive accuracy.

Author 1: Sandra Pajares Centeno

Author 2: Hugo Alatrista-Salas

Keywords: Electricity consumption; forecasting; recurrent neural networks; deep learning