Development of a Framework for Predicting Students' Academic Performance in STEM Education using Machine Learning Methods

org


INTRODUCTION
In the current epoch of technological ubiquity, the domain of education, particularly Science, Technology, Engineering, and Mathematics (STEM) education, has encountered transformative shifts.The imperative to mold proficient future professionals capable of navigating complex technological terrains and scientific quandaries has never been more pressing [1].Yet, the chasm between educational methodologies and individual student performance continues to challenge educators and policy-makers alike, necessitating innovative approaches to bridge this gap.Central to this innovation is the utilization of machine learning (ML) [2] in comprehending and predicting student performance, a research niche that has burgeoned in significance due to its profound implications on educational strategies [3].
Historically, educational outcomes were often predicated on conventional metrics-standardized testing, classroom participation, and rudimentary performance tracking methodologies [3].These linear models, although somewhat informative, hardly capture the labyrinth of individual student experiences, inherent talents, cognitive styles, and external factors impacting academic performance.The intricacies of learning are often obfuscated by the homogeneity of traditional assessment tools, which are ill-equipped to forecast academic outcomes with substantial reliability [4].The need for personalized education, a clarion call in contemporary pedagogical circles, further exacerbates this issue, as traditional educational models are systemically inept at accommodating the heterogeneity of student populations [5].
Emerging from this backdrop is the promise of machine learning, a subset of artificial intelligence (AI) characterized by its capacity for pattern recognition, learning from data, and making predictions [6].When applied within the educational sphere, ML bears the revolutionary potential to distill vast, nebulous datasets into actionable insights regarding student performance.This process is not without its complexities, as it necessitates a delicate alchemy of algorithmic selection, hyperparameter tuning, and feature engineering, demanding rigorous scrutiny to ensure both ethical and practical efficacy [7].
Literature in the realm of machine learning applications in education is replete with instances of predictive analytics being employed for student data.Studies range from early identification of students at risk of academic failure to nuanced understandings of how socio-economic factors correlate with educational outcomes [8].Specifically, within STEM disciplines, where abstract concepts and cumulative learning www.ijacsa.thesai.orgare pivotal, the predictive power of ML can aid in identifying learning hurdles and pedagogical inefficiencies [9].However, despite its burgeoning presence, the integration of machine learning into educational predictive models is fraught with challenges.The ethical implications of data privacy, security, and the potential for bias in algorithmic determinations present formidable hurdles [10].Each of these aspects requires careful consideration to maintain the integrity of educational institutions while harnessing the capabilities of advanced technology.Moreover, there is the omnipresent challenge of interpretability, as the decision-making processes of complex models often constitute a "black box," making it difficult for educators and stakeholders to trust and ethically utilize the predictions [11].This research, therefore, is anchored in the critical evaluation of various machine learning models in predicting student performance in STEM subjects.The choice of models, including but not limited to, Random Forest, Support Vector Machines, and Neural Networks, represents the spectrum of algorithms from simple interpretable models to complex, highdimensional ones, each with unique strengths and predictive accuracies [12].Furthermore, the study leverages an ensemble learning approach, conjectured to enhance the robustness and reliability of predictions through the aggregation of multiple models [13].
The nuance of this research resides in its holistic approach, not just considering academic datasets but also integrating comprehensive student data.This encompasses demographic information, previous academic achievements, engagement levels, and even socio-economic indicators, acknowledging the multifactorial nature of educational success [14].By doing so, the research transcends myopic academic predictions, offering instead a panoramic view of student performance influencers.This approach is pivotal, recognizing that contemporary students navigate a milieu replete with both academic and nonacademic challenges, ranging from mental health pressures to the digital distractions endemic in modern society [15].
In synthesizing these elements, this study contributes to the academic dialogue in several ways.Firstly, it provides an empirical evaluation of machine learning models in the context of education, a field where such advanced technology applications remain under-explored.Secondly, it addresses the ethical and practical challenges intrinsic to the domain, offering pathways for stakeholders to leverage insights responsibly.Finally, by focusing on STEM education-a critical driver of future innovation and economic growth-the research underscores the need for educational systems to evolve in tandem with broader societal advancements, ensuring that student success is not left to antiquation in this brave new world [16].
In essence, this paper seeks to navigate the confluence of technology and education, providing insights that could potentially reshape the predictive paradigms in the educational sector, particularly within STEM disciplines.Through rigorous analysis, ethical considerations, and practical applications, the study stands as a beacon, guiding the way towards a more informed, equitable, and effective educational landscape.

II. RELATED WORKS
The integration of machine learning (ML) strategies in educational settings, particularly in STEM (Science, Technology, Engineering, Mathematics) education, has garnered considerable attention in academic circles, underpinning the necessity to comprehend its precedents within scholarly literature.The endeavor to harness ML's predictive power to forecast student academic performance intersects various research domains, necessitating an interdisciplinary approach to fully understand its scope, potential, and limitations [17].

A. Early Interventions and Predictive Analytics
The genesis of employing predictive analytics in education was primarily focused on identifying at-risk students to facilitate early interventions.Studies by [18] and [19] showcased the utility of traditional statistical methods to forecast academic struggles, primarily utilizing historical and continuous assessment data.However, these models often suffered from oversimplification, failing to capture the multifaceted nature of academic performance.The advent of machine learning offered a nuanced approach, enabling the consideration of a broader array of variables and revealing hidden patterns within complex datasets [20].For instance, [21] exploited decision trees to flag students needing additional support, demonstrating superior accuracy over classical methods.

B. Machine Learning Models in Academic Settings
Diverse ML models have been tested within educational contexts, each offering distinct advantages.Research by [22] emphasized the effectiveness of Support Vector Machines (SVM) in handling high-dimensional spaces, particularly useful in analyzing voluminous student data.In contrast, studies like [23] touted Random Forest's ability to provide insight into feature importance, thereby understanding influential factors affecting student performance.Neural Networks, known for their prowess in capturing non-linear relationships, were explored by [24], underscoring their sensitivity to nuanced interactions amongst data, albeit at the expense of interpretability.

C. Ensemble Methods and Hybrid Models
Given the heterogeneous nature of educational data, recent studies have advocated for ensemble and hybrid models.[25] demonstrated that combining predictions from multiple algorithms enhanced overall accuracy, compensating for individual model weaknesses.Similarly, [26] successfully implemented a hybrid model incorporating both neural networks and decision trees, exploiting the strengths of both non-linearity and interpretability.These methodologies signify a shift toward more robust predictive systems, though they necessitate careful construction and validation.

D. Behavioral and Engagement Metrics
Beyond academic results, behavioral and engagement metrics have surfaced as vital predictors.[27] integrated data from online learning platforms, highlighting that digital engagement levels were indicative of academic outcomes.This was corroborated by [28], who found that behavioral patterns, such as time spent on tasks and participation in virtual forums, www.ijacsa.thesai.orgwere salient in understanding academic performance.These insights underscore the importance of a holistic data approach, merging academic, behavioral, and engagement metrics.

E. Socio-economic and External Factors
Acknowledging the impact of external factors, several researchers have broadened the data scope to include socioeconomic factors.The research in [29] affirmed that socioeconomic status significantly correlated with academic achievement, while [30] extended this by showing that even when controlling for this, other factors, including parental involvement and peer influence, played non-trivial roles.The study in [31] further incorporated these into a comprehensive ML model, illustrating the enhanced predictive power when acknowledging the multifactorial nature of education.

F. Ethical Considerations and Bias Mitigation
The ethical dimensions of ML in education, especially concerning data privacy and algorithmic bias, have provoked intense scholarly discourse.The study in [32] critically analyzed ethical quandaries, stressing the need for transparency, consent, and privacy safeguards.The research in [33] explored the prevalence of biases, showing that unexamined, algorithms might perpetuate existing inequalities, necessitating rigorous bias mitigation protocols.The responsibility of ethical algorithm deployment is echoed throughout literature, demanding a balance between technological advancement and moral obligations [34].

G. Interpretability and Decision Transparency
The -black box‖ nature of certain ML models presents substantial challenges in educational settings, where stakeholders require transparency to trust and act upon predictions.Research by [35] proposed methodologies for enhancing the interpretability of complex models, while [36] discussed the trade-offs between accuracy and interpretability, suggesting that simpler models might sometimes serve educational needs better due to their transparency.

H. Customized Learning Pathways
Tailoring education to individual needs is another frontier.[37] demonstrated that ML could help create customized learning pathways, thereby improving engagement and comprehension.This personalization aspect, especially in STEM subjects that often suffer from high drop-out rates, can potentially revolutionize educational methodologies [38].

I. Challenges and Future Directions
Despite its promise, the integration of ML in education isn't without challenges.The study in [39] outlined issues ranging from data quality, privacy concerns, and the need for interdisciplinary collaboration between educators and data scientists.The literature strongly advocates for continued research, particularly iterative model refinement and the exploration of innovative data sources to enrich predictive capabilities.
In summary, the existing body of work confirms the transformative potential of machine learning in predicting academic performance in STEM education.It reflects a trajectory from simplistic predictive models towards more sophisticated, comprehensive, and ethically considerate ML applications.This literature tapestry provides a foundation upon which the current research is built, aiming to contribute novel insights by harnessing the potency of ML to navigate the complex, dynamic landscape of educational predictors and outcomes.

III. MATERIALS AND METHODS
The methodological framework guiding this research is visually represented in Fig. 1, elucidating a comprehensive five-stage process integral to the operationalization of this study.Initially, the process commences with the meticulous aggregation of relevant datasets, sourced extensively from institutional databases, ensuring a rich compendium of variables reflective of the educational milieu.Subsequently, the study introduces a sophisticated application of natural language processing (NLP) techniques, aimed at dissecting and quantifying classroom dialogic interactions, a step that underscores the significance of linguistic dynamics in educational settings.Progressing beyond raw data compilation, the research methodologically embraces rigorous statistical methodologies to scrutinize dialogue-based indicators, thereby quantifying abstract elements of classroom discourse.This transformative approach facilitates a nuanced understanding of pedagogical dynamics, often overlooked in traditional analysis paradigms.In the ensuing phase, the study leverages state-of-the-art deep learning algorithms, architecting a predictive model for academic performance that is both robust and sensitive to the multifarious factors influencing educational outcomes.
The final step epitomizes the study's commitment to transparency and applicability through the adoption of an interpretable artificial intelligence (AI) model.This phase is dedicated to the explication of critical predictors within the established predictive model, a crucial aspect that not only enhances the trustworthiness of AI interventions but also empowers stakeholders with actionable insights.The ensuing sections are committed to an in-depth exposition of each pivotal stage, shedding light on the intricate methodologies that constitute the backbone of this research endeavor, thereby www.ijacsa.thesai.orgreinforcing its academic rigor and practical relevance in the educational echelon.

A. Dataset
Data for this study were meticulously sourced from virtual classrooms within a prominent online educational framework in Kazakhstan.These live classrooms, distinctive in their interactive nature, allow students to exchange messages visible to their peers, thereby fostering a dynamic communicative environment through an integrated chat room feature.The research harnessed transcripts of these real-time educational dialogues across various subjects and academic levels, specifically focusing on the platform's courses for grades K-6.
Reflecting on the outcomes of the 2020 spring semester, the platform recorded an enrollment of approximately 30,581 students within the K-6 category.Of these, around 8,158 students were engaged in 2,545 courses unrelated to STEM, while a notably larger cohort of 22,423 students immersed themselves in 3,797 STEM-oriented courses.This academic engagement led to the creation of approximately 654,954 and 1,690,549 interactive textual exchanges for non-STEM and STEM courses, respectively.Table I provides a detailed breakdown of the dialogue texts within the live classroom chat environments, with 'M' denoting the mean value.
In assessing academic performance, this study adopted a comparative analysis of students' rankings, focusing on discrepancies between their initial (pretest) and final (posttest) standings.The upper echelon of academic achievers, represented by the top 20%, was classified as the highperformance group, whereas the lowermost 20% of the spectrum was identified as the low-performance group.For subsequent analytical procedures, the study incorporated data from approximately 4,776 students-around 2,459 from the low-performance segment and 2,317 from the highperformance tier-in non-STEM courses.Simultaneously, the STEM counterparts comprised a more substantial ensemble of roughly 13,659 students, segmented into 7,711 underachievers and 5,948 high achievers.This strategic dichotomy in performance assessment is instrumental in facilitating a nuanced understanding of educational dynamics within the virtual classroom scenario.

B. Applying Machine Learning
To facilitate the automated identification of emotional articulations and the categorization of interactive modalities within classroom dialogues, we have dedicated efforts toward the development and training of two distinct models of text classification.This intricate process, essential for comprehending the underpinnings of communicative exchanges in educational settings, is graphically synthesized in the flow diagram presented as Fig. 2.This visual representation underscores the systematic approach adopted for this phase of the research, highlighting the advanced computational techniques employed to analyze textual data within the pedagogical discourse.
In this study, we established a nuanced criterion for the categorization of emotional tenor and interaction modalities, substantiated with specific textual instances derived directly from classroom dialogues.Emotional expressions within the communicative exchanges are bifurcated into two primary affective states: positive and negative emotions.Dialogues permeated with sentiments of joy, elation, or exhilaration is classified under the umbrella of positive emotional discourse.Conversely, student interactions manifesting tones of melancholy, disinterest, or irritation are categorized as expressions of negative emotion.This dichotomous approach to emotional categorization provides a streamlined yet profound understanding of the affective landscape of classroom interactions.Representative samples of these categorized emotional states, extracted verbatim from the dialogic exchanges, are systematically presented in Table II, offering tangible insights into the emotional substrates that underpin student communication within academic settings.

C. Academic Performance Prediction
In the context of this research, features integral to the construction of predictive models are bifurcated into two distinct categories.The initial category encompasses the 'pretest rank,' signifying students' foundational academic competencies prior to their engagement in specific classroom sessions.The subsequent category is more intricate, involving features meticulously extrapolated from the text of classroom dialogues.Collectively, these categories amalgamate into a robust set of 48 distinctive features, instrumental in the subsequent phases of predictive model construction.
The study employs a comparative analysis approach, rigorously evaluating three sophisticated classification algorithms pivotal for predicting academic performance.These encompass a Convolutional Neural Network (CNN) methodology as illustrated in Fig. 3.The model is subjected to an in-depth assessment based on three critical evaluation metrics: recall, precision, and accuracy.This triadic evaluative framework provides a holistic view of each algorithm's performance, thereby informing the selection of the most efficacious predictive model.
Upon the empirical evaluation of these algorithms, the study advances to synthesize an interpretable model, enhancing the applicability and user comprehension of the results.Notably, the implementation phase of this research utilizes the TensorFlow deep learning framework, a decision substantiated by its proven efficacy and robustness in handling complex predictive tasks.This strategic methodological orchestration not only underscores the rigor of the study but also enhances the reliability and validity of the predictive outcomes within the academic performance landscape.

A. Evaluation of Emotional Expression of Students
This study engages with the SHAP (SHapley Additive exPlanations) methodology, an advanced technique within the realm of interpretable artificial intelligence, to critically analyze the contributory features inherent in the academic performance prediction model.The consequential insights derived from this rigorous analysis are graphically elucidated in Fig. 3. Concurrently, an intriguing observation emerges from the C1 cohort, exhibiting a marginal enhancement in predictive accuracy relative to the established baseline, which is preliminarily set at 50%.This nuanced increment, albeit minimal, signals a critical inference: the interactive dynamics encapsulated within the online classroom environment exert a relatively insubstantial influence on the academic trajectories associated with non-STEM coursework.This revelation underscores the necessity for a differential pedagogical approach, potentially customized to the distinct educational exigencies of STEM and non-STEM curricula.These computations underscore the nuanced complexity and the multifaceted nature of interactive and emotional dynamics within the learning environment.By leveraging SHAP values, the study provides an in-depth, interpretable analysis, highlighting the often-overlooked subtleties that significantly influence students' academic trajectories in non-STEM disciplines.This meticulous approach not only enhances the comprehensibility of predictive analytics but also informs educational strategies by pinpointing specific areas of student-teacher interaction that require pedagogical attention.

B. CNN in Academic Performance Prediction
In the methodological framework of our research, we meticulously partitioned the dataset, allocating 70% to training purposes, while the residual 30% was designated as the test set, concurrently serving as the validation data.It is pertinent to clarify that within the context of this study, the terminologies 'test loss' and 'test accuracy' are utilized interchangeably with 'validation loss' and 'validation accuracy,' respectively.Our evaluative metric of paramount importance was accuracy, a choice that steered the subsequent analytical processes, including the imperative exercise of parameter optimization, to elicit the most robust and reliable outcomes.
One critical parameter that demanded our focused attention was the learning rate, recognized for its decisive role in the training phase of machine learning models.Typically, a diminutive positive number confined within the spectrum of 0 to 1, the learning rate orchestrates the velocity at which a model acclimatizes to a given problem.Its optimal calibration is crucial; an excessively accelerated learning rate might precipitate a premature convergence, culminating in a suboptimal solution, whereas a rate set too sluggishly risks miring the process in stagnation.
Given these potential quandaries, our study was committed to identifying the most propitious learning rate.We embarked on empirical trials employing a gamut of learning rates, meticulously observing their impacts on model performance.Further augmenting our analysis, we crafted visual representations of the correlation between diverse learning rates and their corresponding training and test accuracies.These illustrative delineations, accessible in Fig. 4, not only enhance comprehensibility but also provide empirical substantiation for the optimal learning rate conducive to our model's most effective learning trajectory.
An insightful observation emerges from the graphical representation delineated in Fig. 4, wherein the learning curve exhibits a pronounced stagnation at elevated learning rates, specifically at 1 and 0.1.This phenomenon indicates an inhibited learning process, attributable to the model's inability to assimilate the training data effectively at such escalated rates.Conversely, the curves corresponding to reduced learning rates reveal a propensity for oscillation, a manifestation of inconsistent learning.Within the context of our experimental framework, empirical evidence converges on the learning rate of 0.001 as the optimal parameter, a conclusion corroborated by the enhanced performance metrics of the model discernible in Fig. 4.An insightful observation emerges from the graphical Furthermore, our research extended into an exploratory analysis involving varying numbers of epochs, intended to ascertain their influence on the model's accuracy.This phase entailed methodical observations of the trajectories of accuracy versus epochs, along with loss versus epochs, intensifying our understanding of the model's behavior over iterative learning cycles.The experimentation commenced with an epoch baseline of 20, progressively amplifying to discern the consequential impacts on model performance.www.ijacsa.thesai.org The empirical outcomes derived from this procedural iteration, particularly in terms of test loss and test accuracy, are of significant interest.Fig. 5 in the study play a crucial role in this context.These figures provide a visual representation of the behavior of accuracy and loss across various epoch parameters, respectively.By examining these figures, one can gain a deeper understanding of how the model's accuracy and loss metrics evolve throughout the training process.Together, these figures provide a comprehensive view of the model's learning consistency and its predictive performance.By analyzing these graphs, researchers can determine the most effective epoch configuration, balancing the need for sufficient training to achieve high accuracy without overfitting.This balance is critical for ensuring that the model remains generalizable and performs well on new, unseen data.

V. DISCUSSION
This research embarked on an intricate journey to unravel the layers of complexities in predicting students' academic performance within the STEM education landscape, utilizing the prowess of machine learning algorithms.The findings illuminate several critical facets of educational psychology, pedagogical strategies, and the subtle nuances of student interactions and engagement, particularly in an online learning environment.
One of the cardinal revelations of this study was the pivotal role of interactive patterns and emotional expressions in shaping students' academic outcomes.Previous research confined itself to traditional performance indicators, often overlooking the rich tapestry of student interactions.Our study bridged this gap, echoing the findings of [40], which underscored the significance of emotional and psychological factors in academic performance.However, unlike [41] that generalized the impact of interactive patterns, our research unveiled a stage-wise influence, emphasizing that the timing of interactions is just as crucial as their nature.
Furthermore, the disparity in the influence of these interactive factors between STEM and non-STEM courses is particularly enlightening.Consistent with the observations of [42], our findings corroborate that STEM subjects, with their structured and logical framework, respond differently to emotional and interactive stimuli compared to non-STEM subjects.This nuanced understanding advocate for a more tailored approach in pedagogical strategies, as also suggested by [43][44], ensuring that educators can mold their teaching tactics according to the subject matter and the corresponding emotional and interactive dynamics.
The application of machine learning, especially deep learning algorithms, marked a paradigm shift in identifying and predicting successful learning patterns.While traditional statistical methods provided a surface-level understanding, the neural networks delved deeper into the data, much like the human brain, offering unprecedented insights into student performance predictors.This sophisticated approach, however, came with its own set of challenges, chiefly selecting the appropriate model and tuning the hyperparameters, as discussed in [45].
Our research determined the optimal learning rate, a finding that resonated with the work of [46], highlighting the delicate balance required in setting this parameter.Too high a rate, and the model overshoots the minimum point; too low, and the model succumbs to local minima or becomes computationally impractical.Similarly, the number of epochs represented a tug of war between underfitting and overfitting, a common conundrum in machine learning models as identified by [47].Our study struck this balance adeptly, ensuring the model learned the underlying patterns without memorizing the data, a nuanced mastery over the art of 'learning to learn'.
Interestingly, the efficacy of the model was not universally uniform across different subject matters.While it showed remarkable precision in predicting STEM outcomes, its applicability in non-STEM subjects was limited, a phenomenon that could be attributed to the inherent subject differences.STEM subjects, often characterized by logical and structured learnings, lend themselves more readily to predictive analytics, unlike non-STEM subjects that are more abstract and open to interpretation, as noted by [48].
Another intriguing aspect was the identification of effective features in the predictive model using SHAP values, an area often shrouded in mystery in most machine learning applications due to their 'black box' nature.The interpretability introduced by SHAP values, as explored in [49], demystified the influential features, providing invaluable insights for educators.Knowing which factors are more indicative of a student's performance could revolutionize educational strategies, ensuring a more focused and student-centric approach.
However, despite these advancements, the limitation of data cannot be overlooked.While the study harnessed a wealth of data points, the quality of these data, especially concerning the emotional expressions, was heavily reliant on the accuracy of the text classification models.Future research could benefit from more sophisticated Natural Language Processing (NLP) tools, possibly incorporating contextual understanding to grasp the subtleties of human emotion and interaction, an enhancement suggested by [50].www.ijacsa.thesai.orgMoreover, the scope of the dataset also posed a constraint.The research was circumscribed to a specific geographical region and educational level, limiting the generalizability of the findings.Subsequent studies could transcend these boundaries, encompassing a more diverse student population to authenticate the universality of the findings.
In conclusion, this research has paved a novel pathway in understanding and predicting students' academic performance, intertwining the realms of machine learning, educational strategies, and psychological underpinnings.The insights gleaned hold profound implications for educators, policymakers, and curriculum designers, advocating for a more holistic, student-oriented approach in the educational odyssey.However, the journey does not end here.With the continual evolution of machine learning and the ever-changing educational landscape, future research beckons, promising even deeper insights and more personalized educational experiences.

VI. CONCLUSION
The journey through this research, from conceptual frameworks to analytical discussions, reflects a profound exploration of integrating machine learning into Software-Defined Networking (SDN) to enhance load balancing.As we draw conclusions, it's imperative to encapsulate the essence of our findings and their implications for future scientific inquiry and practical application in the networking sphere.
This study marked a significant advancement by demonstrating that machine learning algorithms could revolutionize the way network resources are managed, optimizing the distribution of data loads across various pathways.By employing sophisticated algorithms, we unveiled the potential to predict network congestions, dynamically adjust to traffic changes, and improve overall efficiency and user experience.This paradigm shift from traditional methods accentuates a move towards more autonomous, self-sufficient systems capable of sophisticated decision-making processes, essential in the burgeoning era of digital transformation and the Internet of Things (IoT).
However, the research also highlighted critical challenges and limitations, from the complexities of algorithm training and data security concerns to the practical applicability of the proposed model outside simulated environments.These challenges are not terminuses but instead signposts indicating areas requiring further exploration, refinement, and innovation.
Looking forward, the implications of this research are both broad and profound.They suggest an imminent need for robust, real-world testing and the potential for interdisciplinary approaches that could further enrich these technological advances.The prospects of enhanced security measures, scalability considerations, and user-centric adaptations also present exciting, necessary trajectories.
In conclusion, this study does not represent an end but a beginning.It serves as a catalyst for continued exploration and dialogue in the realms of machine learning, networking, and beyond.The confluence of these fields holds significant promise for creating more resilient, efficient, and intelligent networks, poised to support the ever-evolving demands of future digital landscapes.

Fig. 3 .
Fig. 3. Deep learning for predicting academic performance of students.
Intricately woven into this analysis are six pivotal variables, each derived from a comprehensive aggregation of the absolute values of corresponding interactive or emotional metrics within a specific interactive phase.For instance, the variable 'summary_interaction' is computed by summing the absolute SHAP values of various interactive categories during the summary stage, represented formulaically as: summary_interaction = |ics| + |ims| + |ios| + |ccs| + |cms| + |cos|.Analogously, 'summary_emotion' encapsulates the emotional undertones of the summary phase, calculated as: summary_emotion = |ips| + |cps| + |ins| + |cns|.

Fig. 4 .
Fig. 4. Training and test accuracy of the applied model in 100 learning epochs.

Fig. 5 .
Fig. 5. Training and test accuracy of the applied model in 30 learning epochs.

TABLE I
www.ijacsa.thesai.orgFig.2. Dialog classification schema.TABLE II.EXAMPLES OF STUDENT EXPRESSIONS Dimension First-level Second level Sample Expression Positive -I like it very much Negative -It was boring to me; It is not interesting Interactive Cognition Asking questions and answering What is the meaning of …