Distributed Focused Web Crawling for Context Aware Recommender System using Machine Learning and Text Mining Algorithms

— In today’s world, Recommender System (RS) is the most effective means used to manage the huge amount of multimedia content available on the internet. RS learns the user preferences and relationships among the users and items. It helps the users to discover new interesting items and make use of different media types such as text, audio, video and images. RS can act as an information filtering model which can overcome the issues related to over-fitting and excess information. In this work, a new distributed framework named DAE-SR (Deep AutoEncoder based Softmax Regression) is introduced for context-aware recommender systems which focus on user-item based interaction and offers personalized recommendations. The proposed model is implemented in PYTHON platform. The dataset used for experimentation is Foursquare dataset. The performance of the proposed context-aware RS is beneficial to both the users and service providers. Its helps in decision making process and can offer relevant recommendations to users. The performance is evaluated in terms of various metrics such as accuracy, recall, precision and so on. From the implementation outcomes, the proposed strategy achieved good accuracy (98.33%), precision (98%), run time (1.43 ms) and recall (98.1%). Thus, it is proved that the proposed DAE-SR classifier performs better compared to other models and offer dependable and relevant recommendations to users.

RS aims to seek the satisfaction of the individual/user via personalization and chooses the most appropriate services, items and provides it to the users considering the details such as user profile, opinions, preferences, purchase history, relationships with clients and interactions with services and products offered. The dual main entities of the RSs are the users, items which learn to create predictions by learning the parameters which minimize the loss among the actual and predicted preferences. Moreover, the RSs can be classified into 3 types such as CF (Collaborative Filtering), CB (Content-based) and HR (Hybrid Recommendation). Among these three categories, CB model gain advantages over the traditional RS system by exploiting the features of contextinformation as user activity, time, and location.
The integration of RS with context information influences the value of these models by enhancing the possible relevant recommendations in accordance to the varying user requirements. The CEP (Complex Event Processing) is integrated with RSs to provide personalized recommendations and allows the business users to put on event-driven rules for controlling and filtering the input parameters of recommendation engines [12][13][14]. The concept of contextbased RSs minimizes the gap among the information models and users, therefore these information models can actively realize the user's context and offer better experience to users [15][16]. Recommender systems (RSs) are a subclass of data (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 14, No. 3, 2023 602 | P a g e www.ijacsa.thesai.org filtering frameworks that look to anticipate the rating that a client would provide for an item.
Recommender frameworks have been connected to variety of uses, e.g., films, music, news, books, look into articles, seek inquiries, social labels, and money related administrations. As a rule there are three different ways to plan recommender frameworks, i.e., cooperative separating, content-based filtering [17], and the hybrid filtering [18]. Recommender models have attracted consideration in both industry and academia. Such frameworks help to oversee data over-burden via assembling the data self-sufficiently and fitting it to singular interests [19]. A recommender framework would have the capacity to catch the clients' preferences. Social Recommender Systems (SRSs) or Recommenders for the Social Web demonstrate the client's preferences by utilizing the data he or she and their companions have distributed in online informal communities [20][21].

A. Motivation
Recommender Systems gained benefits by solving the problem of handling abundant data in the internet. Individuals plan to make effective and smart decisions on the huge number of available choices in various fields such as: a place the person wishes to visit, a movie the person likes to watch, a book he/she wants to read etc. The main motive behind this generation of recommender system is to offer the intelligence that benefits both the consumers and organizations. However, there is an emerging trend of using contextual information to afford meaningful personalized recommendations. The context enhances the basic interaction of user-item to realize recommendations of higher quality. Moreover, recommendation using content-based framework effectively use the information content describing the item or user such as item description or user profile to execute accurate recommendations. Also, content-based (CB) recommenders are more suitable with sparse datasets and may not suffer from the cold-start issue. As the textual data is used to represent the item description and user profile, this content based contextaware model make use of the common evaluation metrics to measure the similarity between the contents. The recommendations based on contextual information emphasize that the choice made by individuals is liable on certain context, rather than being invariant of it. Sometimes, the same users might select diverse services/products under a dissimilar context. Hence, this condition motivated the authors to develop a precise prediction of preferences by consumer depending on the appropriate contextual data which is being combined with the recommendation method.

B. Contributions
In this work, context-aware approach is developed by utilizing the contextual user information, in order to produce more accurate recommendation according to user's preferences. This recommendation model makes use of the information from the dataset in order to create user recommendations on basis of given number of choices. The main novelty of the proposed DAE-SR recommendation model is to develop a context-aware/content-based RS that estimates the characteristics of restaurant and relevant context factors against the user preference model to create top recommendations for restaurant names based on the similarity score calculated. These scores can be used later to sort out the top restaurants for each user according to their expectation. Our proposed (DAE-SR) RS adopts a user preference model by using features of liked/visited restaurants by the user and takes relevant context factors, e.g., restaurant ratings, location, current weather, and time of the day. Most of the existing research works use CF based recommendation because of its higher performance accuracy. Since, CF approach is more effective and reliable one of the major drawbacks reported was the higher scalability. Hence, a new recommendation framework is presented in this work based on context-aware which improves the quality of recommendations. The major contributions of this paper are summarized as follows:  To introduce a distributed framework for context-aware RS based on text mining and machine learning algorithms in order to afford quick and correct suggestions to user on good restaurants name.
 To present the content-based context-aware RS to create an item list that reveals the user's preference and minimizes the error in the classification process.
 To develop a hybrid DAE-SR based recommendation on restaurant names which is trained with the metaheuristic algorithm called JO (Jaya Optimization) algorithm which has good ability of exploration and exploitation.
 The three similarity measures such as Dice's, Jaccard and Cosine similarity coefficients are used to compute the similarity of the content and to obtain high accuracy in receiving top recommendations.

C. Organization
The paper is organized as follows. In Section I, a brief introduction about the topic is given. Section II gives the works related to our proposed algorithm and Section III gives the proposed methodology. The result of the proposed algorithm is provided in Section IV and the entire work is concluded in Section V with future suggestion and references.

II. RELATED WORKS
The main motivation of related works section is to explain about the importance behind each research. In this this section, the discussion about the features of existing approaches, its applications, limitations, metrics and datasets used is clearly described in Table form. Afolabi et al. [22] have presented a semantic web content mining scheme for RS in shopping by online. Through a web crawler, the web textual dataset was gathered built in Java. The combination of an existing ontology and a developed ontology was utilized for the textual data's semantic preprocessing. Next, the recommendation was created by the Naïve Bayes algorithm. The large dimensions of attributes were handled by the annotations to ensure the preprocessing models optimized semantic understanding.
Amato et al. [23] presented a technique SOS: A multimedia RS for Online Social systems. In this viewpoint, the RS was introduced for supporting the browsing of (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 14, No. 3, 2023 603 | P a g e www.ijacsa.thesai.org information accumulations and helping users to discover what they truly required. A new RS for big data applications was introduced in this research, which was able to deliver recommendations based on the interactions between users.
A hybrid learning based recommender framework was proposed by Tarus et al. [24]. In this paper, hybrid information based RS was introduced on the basis of sequential pattern mining (SPM) and ontology. The domain knowledge about the learning and learner resources was represented and modelled by the ontology, although the SPM algorithm determines the learners' sequential learning patterns.
RS with Linked Open Data (RS-LOD) and matrix factorization process (MF-LOD) was proposed by Natarajan et al. [25] for reducing the problems of cold-start and sparsity. The cold start issue was solved by the RS-LOD approach, and the data sparsity problem was eliminated by the MF-LOD approach. The hidden feedback data was elongated the user vector, and the similar semantic items were elongated each item. The collaborative filtering RS's precision was enhanced by the Semantic features from LOD.
Iqbal et al. [26] proposed a context-aware kernel-based recommendation scheme that creates the design over the useritem matrix of context rating. The importance of contextual information was considered for this recommendation. The KCR (Kernel Context Recommendation) algorithm was accurate, flexible and scalable enough to include the various contexts and was utilized for creating the practical recommendations.
A Preference Relation based PMF framework for RSs was proposed by Pujahari and Sisodia [27]. User's PR was taken as input, and the recommendation was also generated. The user's and item's neighborhood data were integrated into the design. The users' preferences towards items were obtained by the Probabilistic Matrix Factorization (PMF) system. This approach was introduced for Collaborative Filtering (CF) in RS. Aghdam [28] proposed a hierarchical HMM (hidden Markov model) in which the changes were identified in the preferences of user on time through modelling the user's latent context. The user as a hidden Markov process was modelled by the user-selected items.
Sangaiah, A.K et al. [30] developed a new technique with the help of ML (Machine Learning) methods to conserve the confidentiality of PBSs (Position Based Services) roaming users. The emergence of the three-phase procedure was used for the identification of PBS user position by means of combining k-NN (k-Nearest Neighbours) and DT (Decision Trees). The destination of the user was estimated by means of the HMM (Hidden Markov Models) and the position track sequence. The service policy followed was the mobile edge computing which was useful for ensuring the PBS timely delivery.
Sangaiah, A.K et al. [31] presented an energy-aware model (EAM) named green adversary for the application of smart industrial environment. The proposed model conserves the information and position confidentiality whereas the existing models make use of the cyber-physical systems by mutually using the software and hardware parts in which the energy consumption was minimized. However, this limitation was tested on the recommended approach related to the CPS (Cyber Physical Security) smart industrial applications and has removed the severe effects and also tested on CPS realworld applications.
Sangaiah, A.K et al. [32] presented whale optimization algorithm (WOA) to obtain optimal solution for network issues. WOA was powerful heuristic algorithms that boosted the exploration process and prevented bottleneck. However, the main drawback of this algorithm was a low convergence rate. Likewise, Bat algorithm (BA) was a novel meta-heuristic algorithm presented by Sangaiah, A.K et al. [33] to achieve optimal solutions in terms of sustainability and quality. It has the ability to monitor and control the network activities. Though, the performance of BA for huge data was not efficient. It was considered as a major disadvantage in BA.
Gupta Garima and Rahul Katarya [36] presented an Ensemble Particle Swarm Optimization (EnPSO) approach which was considered as an AutoML (AutomatedML) to minimize the model selection complexity. The presented EnPSO optimizes the recommendations intelligently using the identification of best ensemble model. The dataset used for testing was MovieLens in which the recommendations were improved with the AutoML system. The accuracy of the model was verified with the RMSE metric with lower error as 0.918 compared to the baseline IBCF (Item-Based Collaborative Filtering) with Singular Value Decomposition (SVD) having 0.961 higher errors. The main challenge was to find an effective way to reach the performing scheme with the vast search space. The limitation of the suggested approach was that EnPSO cannot provide effective performance gains within the constrained time for some cases.
Katarya Rahul, and Yamini Arora [37] developed a novel scheme called CapsMF (Capsule Networks Matrix Factorization) for product RS. The DNN (Deep Neural Network) text analysis architecture was enhanced with the Bi-RNN (Bi-directional Recurrent Neural Network) for the representation of text descriptions in robust form. The integration of Probabilistic MF with the DNN has the ability to generate recommendations in improved manner. The accuracy of the system was analysed with MAE (0.8878) and RMSE (1.157) shows improved outcomes. The major limitation observed was the training time was quite higher with the Caps Network during the experimentation process.
Katarya Rahul [38] presented a reliable RS model using the improved CF technique. A new self-devised method named P-distance algorithm was used to predict the ratings and acts as a dual-level filtering procedure to identify the nearest neighbours. The experiments were tested on 3 datasets such as Douban, Movielens and Jester. The performance was evaluated with MAE and RMSE metrics to show the effectiveness of the proposed algorithm. Sparsity was the major problem identified with this reliable RS.
Katarya Rahul [39] developed a hybrid RS utilizing KMC (K-means clustering) with ABC (Artificial Bee Colony) optimization procedure. The hybrid KM-ABC approach offers accurate movie predictions by considering the user ratings.
(IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 14, No. 3, 2023 604 | P a g e www.ijacsa.thesai.org Various metrics such as accuracy, recall, precision, MAE were evaluated for verifying the system performance on the Movielens dataset. The experimental results on the dataset indicate that the proposed KM-ABC based recommender system has gained improved performance regarding reliability, accuracy, and personalization for the recommendation of movies. The drawback associated with this system was that only user ratings were considered and the other user characteristics were not considered.
Yadav et al. [40] presented a review on RS using FL (fuzzy logic). With the combination, with fuzzy logic three categories were reviewed such as content-based, memorybased and model-based systems. FL was utilized to assign the fuzzy values to the user items and represents the preference of user in robust form. The Fuzzy recommender was used to compute similarity using the fuzzy values rather than the userbased crisp data. This review offers significant knowledge about the various fuzzy based RS models. The fuzzy logic imitates the decision making process of humans which cover all the intermediate values.
Katarya Rahul [41] developed an improved RS utilizing dual-level MF (matrix factorization). This improved framework incorporates product ontologies for learning the items recommendations. The proposed matrixfactorization based RS initiates each item by generating similarity measures with each other. The similarity was created with the help of genres and a matrix was created using the given user ratings, so that the user gets his corresponding recommendation. Sparsity was the major problem identified with this reliable RS. The dataset used for experimentation was MovieLens dataset and the metrics used for evaluation were MAE and RMSE.
Katarya Rahul and Om Prakash Verma [42] introduced a web-based RS which was mainly based on successive information of user's navigation on web pages. FCM (Fuzzy C Means) approach was integrated to receive top-N clusters. Weights were calculated for each of the page category and the top page recommendations were predicted for the targeted user. The experimentation was performed on the real-world MSNBC dataset. Compared to other clustering approaches the proposed FCM model has gained 33%accuracy. The proposed RS required lower computation speed of 402 seconds whereas the other methods show similar as well as enhanced outcomes.
Katarya Rahul and Om Prakash Verma [43] presented a hybrid model for music recommendations. The proposed HMRS (Hybrid Music RS) represents the combination of both collaborative-content features. The algorithms such as BF (Bellman-Ford) and DFS (Depth First Search) were utilized with the multi-layered context graph. Here the similarity score was computed using PC (Pearson Correlation) coefficient. The overall results were optimized with the PSO algorithmic procedure. The music dataset used was Last.fm and the accuracy of the system was evaluated with recall metric on Top-N recommendations.
Katarya Rahul et al. [44] developed a location-based RS framework with the help of improved RW (Random Walk) model. Both the structural as well the attribute properties were utilized to recommend the location in SNs (Social Networks).
The computations of accurate recommendations on locations were done with the improved RW model. The experimentation was conducted with three real-world datasets such as Gowalla, Brightkite and Foursquare. The accurateness of the system was evaluated with precision, F1 score and recall metrics.
Katarya Rahul et al. [45] presented a HF (Hybrid Filtering) approach in analysing the user behaviour in content-aware RS. The dual filtering strategies used are the pre-filtering and postfiltering models. The HF model eliminates the recommended items which has the lesser chance of relevance. Based on the significance of the contextual attribute both the filters were combined which forms a hybrid model. This HF approach minimizes the sparsity problem also it was quicker than the post-filtering method. The experimentation was performed on 50 movies that were selected equally from Movielens dataset and the precision calculation determines the accurateness of the system. Sanchez et al. [46] presented a RS for food delivery on the basis of number of orders. Nearest-neighbour (NN) algorithm was used for the evaluation of users preferred restaurants and buying patterns. Teixeira et al. [47] introduced a diabetic friendly restaurant recommendation using MAS (Multi-Agent System) for multi-criteria decision making. The users with diabetic illness can be benefitted and can find out suitable nearby restaurants. Table I represents the review on various recommendation systems.

A. Research Gaps
Most of the works developed in the context aware recommender system concentrates only on improving the accuracy of recommendation. But, the major purpose is to improve the recommendation accuracy when considering the design objectives such as a user's item's context. Moreover, the main challenge of a RS is to generate a significant recommendations using context based rating of user and item information. Hence, the recommender system should be accurate, flexible and that should predict the contexts accurately. Moreover, some of the recommender systems are providing the improved accuracy but not providing the rating of contexts. Some of the recommendation systems shows decreased computational complexity but fails in the accurate prediction and also results in the difficulty of generating recommendations.
The running time of the previous RS is noted to be higher while providing the recommendations which depend on the previous preferences of users. The preference may sometimes change based on time because of the mood change of users otherwise the change of contexts. Numerous earliest techniques are presented based on the statement of every user having the static pattern. But without these changes, the recommendation does not equal the user's choices. The context aware dependent recommendation systems are dealing with this problem through the contextual data. The contextual data usage is difficult because of attaining all contextual data for processing. Moreover, the different types of contexts used in the RS extend its dimensionality. Hence, to overcome all these issues, the authors of this work presented a Distributed Focused Web crawling with Machine Learning and Text Mining algorithms for the effective context aware recommendation system.

B. Research Questions
The recommended system aims to execute certain tasks in assisting smart recommendations on restaurants. In this work, context aware-RS have played a most significant role in generating recommendations. Thus, some of the following research questions presented are:  How are the contexts collected?
 Which contexts are used in smart learning?
 What data mining and recommendation techniques used to processes the contexts?
 What are the recommended activities?
The existing approaches utilized for the recommendation system was discussed in Section II with its research gaps. To avoid the issues presented in the existing RS, a new approach (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 14, No. 3, 2023 606 | P a g e www.ijacsa.thesai.org is proposed and the research questions are framed which is solved in this presented work (Section III).

III.
PROPOSED METHODOLOGY The recommendation system uses one of the exceptional pieces of the researches, which can be used to offer better solution for challenges also it uses supplementary refined machine learning algorithms for example, ANN. To analyse the performance of the Context-aware recommendation systems a combination of different approaches is used. The initial attempt is to model the criteria rating, approximately powerful machine learning approaches are used for this process. In order to enhance the prediction accuracy the deep learning model is used in this work. This paper make use of deep auto encoder and Softmax regression trained (DAE-SR) with a Jaya optimization (JO) algorithm for providing effective recommendations. We selected the Jaya optimization algorithm, since it has a high tendency to high convergence rates, low training time, and escape local minima compared to other optimization algorithms. Fig. 1 shows the overall workflow diagram for the proposed strategy. For the web data acquisition, crawler is a significant tool. However the frequent updates of the data sources, distribution channels and web data structures are resulted in high costs of crawler program maintenance and development. From the raw web data the textual information's keywords are extracted by means of measuring keyword comprehensive weights via informative features. Textual information's keywords reflects the ideal way for data mining also establishes the knowledge representation model, after that constructs the index library. At last, for the deep machine learning strategies build a prediction model to obtain relevant recommendations. The prediction is performed based on the strategy named as DAE-SR to obtain effective results. The textual information pre-processing consists of following phases. They are:  Information Collection and Processing.

A. Information Collection and Processing
For collecting the news the web crawler provide important theoretical support. Web crawler is a central part of search engines. The use of web crawler is not difficult when getting the news from the internet. Implementing a new crawler and not extending the existing crawler module by the user is the major problem with crawler. It is used to rescue the web pages also attach them into the local source. In a centralized location, collecting and processing the entire contents of web is the general-purpose of web crawlers. Every day a huge volume of web pages are frequently added and data is persistently changing due to the rapid emergence of internet. On the basis of event pattern or event procedure rule, the CEP is said to be a real-time data processing methodology. It is used to retrieve the high level knowledge from the large extent of data. It can also be utilized to analyze the trends, track the data from multiple sources, and patterns. The events that happened in past, are used in any order which can be allowed by CEP.

B. Textual Information Pre-processing
Pre-processing strategies named as stop words removal, word extraction, stemming and also term frequency-inverse document frequency (TF/IDF) are the important approaches in textual information pre-processing. Based on the words existing in the user opinion feedback reviews, the words are tokenized, count of frequency of word and stem are done by using the pre-processing strategies. Fig. 2 shows the process of pre-processing.  Pre-processing plays a vital role in text mining (TM) applications. Pre-processing is the first step, which involves the combination of other steps such as token extraction, stemming, removal of stop words, and TF/IDF tasks.

1) Word (tokenization) extraction:
The procedure of converting the sensitive data into tokens is called as tokenization. Here the significant amount of text structure is reduced to a word, phrase, or symbol. Using NLP (Natural Language Processing), the word in the document set D is tokenized.
2) Stop words: Pronouns, prepositions and articles are generally present in documents. These words may not add much meaning and considered as useless. Some of the stop words used are ‗with, an, is, the, has, in, be and, at' etc. In text mining applications stop words are not considered as keyword, so it is eliminated from the document.
3) Stemming: The process of extracting the base form of words by eliminating the affixes and stem the root of a word into the original word is termed as stemming, example; programs, programming, programmed  ‗program' etc.  1 are correspondingly, based on the used term weighting system. Table II gives the example of two dimensional based VSM model. b) Term frequency (TF): In the web page, TF is related through the weight of a definite term. Normalized TF is measured in this paper. By finding the amount of raw frequency of the term t in web page w is calculate weight of term t is displayed in Eq.  Table III shows the example of collection of web pages, in this highest TF value is mentioned as bold and underlined cells is the second highest value in the matrix. As shown in Table III, on page 3 the term ‗payment' has the highest TF value (0.7071). The term ‗payment' is placed two times in page 3 and 4. The TF value is normalized to webpage's Euclidean norm. * The number of distinctive terms in the web page ** Summation of term occurrences in entire web pages. *** The number of web pages in which the term appears. 608 | P a g e www.ijacsa.thesai.org c) Document frequency: The number of web pages within the collection is represented as DF (weight of term t), here the term t is found. It is a global weighting term.

4) Term
TF-IDF: The terms within the collection of web pages are denoted by TF-IDF, which is an effective ranking measure. It also reproduces the statement that less frequent word in the collection is more significant word in web page and viz. The dot product of TF and IDF of a term is denoted as TF-IDF. Table IV gives the example on the basis of TF-IDF weight scheme.

DF N IDF
The term ‗payment' appears three times in the document, so it has highest TF-IDF value shown in Table IV. Also, the term ‗banking' and ‗service' appears twice in the collection.

C. Similarity Calculation
After the text pre-processing the opinions are grouped together according to their similarities, while the huge number of reviews are received. Similarity measurement is necessary in text mining approaches. Similarities among the string are calculated by similarity measures. Dice's coefficient, Jaccard similarity coefficient and cosine similarity are defined below:

1) Dice's coefficient:
This metric is utilized to measure the documents similarity, and the inventive formula is given in Eq. (4).
Where, |B| and |C| signifies the number of terms present in documents. SQ represents the similarity quotient. 3) Jaccard Similarity (JS) Coefficient: JS is utilized to match the diversity and similarity of model sets. The other names are Jaccard coefficient/ index. The size of intersection divided by the size of union of the sample sets is known as Jaccard coefficient and computes the similarity among finite set of samples as shown in Eq. (7), Where, B and C represent the dual sets.

D. Recommendation Prediction: Hybrid DAE-SR Model
The performance of proposed recommendation model is investigated via neural systems to analyse the textual item content. In this paper, we proposed Softmax regression-based deep auto encoder network (DAE-SR) model trained with Jaya Optimization (JO) algorithm. Compared to the existing models activated in RS, in which the DAE provides better outcomes because of its high capability to reconstruct the inputs. Meanwhile, the softmax regression is utilized for classifying the ratings. We picked the JO algorithm since it has a high propensity to escape nearby minima, low preparing occasions, high assembly rates contrasted with other enhancement calculations, which is utilized to produce improved recommendation prediction.   Fig. 3 represents the architecture of DAE-SR model. It can be seen that DAE and softmax regression can be perfectly linked as a soft-hybrid model, which can make full use of labelled instances and achieve supreme accuracy with Jaya optimization algorithm. Here, we mainly explain the hybrid model of DAE-SR for recommended prediction. The deep auto encoder which is the popular deep learning model can learn discriminative features to produce autonomous distributed depictions of contents of items and users. Table V signifies the details of hyperparameters. A DAE is said to be a deep neural network, which is fabricated via stacked Auto encoders. Pre-training that stacked AE from the bottom to top is a major phase of training a DAE.
The Softmax function computes the probability that the training sample x i belongs to class j given the weight and net input z i . The process of finding the optimal weight parameters is the need for training DAE which minimize the objective function, where the models are trained by JO algorithm [29]. Compared to other optimization algorithms, this JO algorithm is said to be a powerful global optimization approach which can be used to solve the constraint as well as unconstrained problems using the benchmark function. Some of the advantages of JO algorithm are: It does not have any specific algorithm parameters to conduct the real computational experiments and has better algorithm convergence. The supreme reason to choose JO is the victorious nature of this algorithm which makes it more powerful than other evolutionary and swarm intelligence algorithms. The main objective of JO algorithm is, for a particular problem if the solution is once achieved the optimal result is reached simultaneously thereby avoiding the worst result. JO algorithm is most pretentious in application viewpoint and endeavours extraordinary accomplishments to discover the genuine solution successfully. Compared to other optimization procedures this JO algorithm utilizes only dual parameters such as number of iterations and size of population. An imperative benefit of using this algorithm is that it minimizes the time required for the optimization process and ignores the endeavour of varying constraints.  restaurant names by considering the user reviews. Here, DAE-SR is trained on a data. This network gains knowledge from this data, which is compiled as -weights‖ of the network. These weights optimized with the JO algorithm have improved the performance of DAE-SR network in recommendation system. This optimization has good exploration and exploitation ability and hence avoid local optima problem with good convergence rate. The method presented in this paper gives a new way of improving the accuracy of the system.  Fig. 4. Flowchart of proposed model. Fig. 4 displays the flowchart of proposed model. The process starts by collecting the data from the Foursquare dataset. Next step is pre-processing of textual information with the stemming, tokenization, stop words removal and TF-IDF approaches. Three different similarity measures such as dice, Jaccard and cosine metrics are used to evaluate the similarity among the features. Among the three measures, cosine similarity has gained improved performance which can be used for further processing. Finally, the recommendation prediction is performed with the DAE-SR machine learning model. Here, the proposed DAE-SR model is optimized using JO algorithm that generates optimal best solutions with the feedback model and the Top recommendation on restaurant names is displayed. Some of the advantages of the proposed work are:  Savings in cost and processing time.
 No cold-start issue and work well with sparse datasets.
 Quicker than traditional recommendation systems.
 Effective decision making and.
 Useful for people who visits new places.

E. Answers to the Research Questions
The presented context-aware recommendation model extracts valuable information from learner contexts and also leads to obtain better recommendations. In this wok, the contexts were collected using Search engine (Foursquare dataset). Next, the proposed RS named (DAE-SR) adopts a user preference model by using features, e.g., restaurant ratings, location, current weather, and time of the day. The recommendation technique used to process the context is DAE-SR with JO algorithm. The main recommended activity of the proposed model is to generate relevant and personalized recommendation about restaurants.

F. Application of the Proposed Work
Restaurant Recommendation is very useful to users because users can get information easily about different restaurants worldwide. These recommendations offer information's like popular restaurant names, menus, location, phone numbers and directions at a glance via the user's (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 14, No. 3, 2023 612 | P a g e www.ijacsa.thesai.org smartphone. Nowadays, there are number of social networking sites in which users post reviews about restaurants and rate businesses. However, before the creation of these online sites people went to restaurants occasionally visiting an interesting place or hearing reviews from neighbours, friends etc. With the technological advancements, there are number of online websites in which the people can check the details about the restaurant before visiting it directly. The information can be available in the form of ratings, price ranges, reviews, and hours of operation. After analysing all these factors, users can make valuable decisions. Checking the information via these online sites seem to be a time consuming task since users have to visit number of pages about different recommendations in order to find the right restaurant of their choice. Above all, when people visit a new place, travel or move to new place users prefer the same search procedure again. Therefore, the proposed DAE-SR model offers an effective recommendation system for the selection of restaurants. Using this we can build a customized real-time restaurant recommendation system. With this proposed system, we can provide recommendation for restaurants that will suit people's preference. For example, consider a newcomer who has just visited Chennai. He wants to have 'pricey but good selection of beer snobs'. He is unsure about the good restaurants around his location. By inputting, this search query into our proposed model, he can get the top restaurants nearer to his location that are highly rated and get him interested in using our recommendation system as a user. Fig. 11 shows the best example as the application of the work in real world. Experimental results have shown that the proposed DAE-SR model achieved better performance than the existing recommendation procedures.

IV.
EXPERIMENTAL RESULTS The recommender system helps the users in an effective manner to obtain useful and personalized information, in order to make comprehensive decisions in the day to day life of consumers. In experimental analysis, enactment of proposed DAE-SR is associated with the existing classifiers for example, K-Nearest Neighbour (KNN), Deep Neural Network (DNN) and Artificial Neural Network (ANN). For evaluating the proposed approach, the data is being collected from the public source and the dataset is named as Foursquare dataset. The restaurant reviews are discussed in this dataset. It contains restaurant name, location, day and time and the reviews. In terms of accuracy, precision and recall the proposed strategy is estimated. The implementation tool used in this work is PYTHON. The performance is compared based on the similarity. Dice's coefficient, Cosine similarity and Jaccard Similarity Coefficient are the tree similarity methods, which are used in the proposed strategy. Cosine similarity accuracy is better than the other similarity.

A. Description about Dataset
The dataset used for the implementation process is the Foursquare dataset. This is a publicly available data source being collected from https://www.kaggle.com/danofer/ foursquare-nyc-rest?select=readme.txt. However, this dataset covers the details of check-in, tag and tip data of the restaurant locations in NYC gathered from October 24, 2011 to February 20, 2012. The foursquare dataset comprises of 3,112 users, 3,298 venues (locations) with 27,149 check-ins and 10,377 tips. Customer satisfaction can be analysed based on the reviews of reviewers acquired by the restaurants. Hence, the successful growth of restaurants is mainly based on the reviewer's reviews. It contains the details such as restaurant name, location, day and time and the reviews. Here, 70% of data is used for training and remaining 30% of data is used for testing purpose.

B. Evaluation Metrics
The procedure of analysing, collecting and reporting the data regarding the system performance is termed as performance evaluation. The performance of machine learning models is measured by means of the suitable evaluation metrics such as accuracy, recall and precision.

C. Performance Analysis
The analysis of proposed DAE-SR is shown and compared with DNN, ANN, KNN for Dice's Coefficient, Jaccard similarity and cosine similarity are explained in this section.       Fig. 7 displays the analysis of DNN, ANN, KNN and proposed DAE-SR classifier on cosine similarity. The accuracy acquired with DNN, ANN, KNN and DAE-SR is 89, 93, 96 and 98.33% respectively. From the above results, it is finely known compared with other classifier the proposed DAE-SR approach achieved better performance. Also, cosine similarity based performance is high when compared to the other similarity. As in our proposed method, the generating value accuracy is better in Cosine similarity. Table X represents the comparative results of RMSE and MAE. The proposed DAE-SR model shows lower error values which clearly signify the accurateness of the system compared to the existing methods. Minimum or decreased in RMSE and MAE is considered as better outcome. Error minimizing gives better accuracy for the proposed scheme compared to other approaches. In figure 8, compared to two existing approaches, the proposed strategy produced minimum error outcomes.    Table XI gives the various performance outcomes comparison regarding different existing works. Fig. 9 shows the precision, recall, and accuracy comparison of DAR-SR with different existing works. In this research, we have used DAE-SR for the effective recommended system. The proposed approach produced better outcomes than the other two previous works. Our proposed model achieves maximum 98% precision, 98.1% recall and 98.33% accuracy values. The proposed scheme achieves above 98% performances because of the best classification approach DAE-SR with JO.  Table XII signifies the accuracy comparison with existing methods. The accuracy of the proposed DAE-SR model is compared with various state-of-the-art models to evaluate the performance of the proposed recommendation procedure. The existing models on recommendation system gained lower classification accuracy with RS-LOD, RNN, Ensemble, LSTM, GRU and Bi-LSTM as 84.1%, 86.00%, 92.36%, 89.00%, 90.00% and 89.00% respectively. The proposed model achieved supreme accuracy of (98.33%) because the context-aware RS based on text mining and machine learning algorithms afford quick and correct suggestions to user on good restaurants name. Table XIII specifies the run time comparison with existing methods. From the table it is observed that the proposed DAE-SR recommendation model requires less run time 1.43 milliseconds for computation whereas the other models specify higher computational time which reduces the system efficacy.    615 | P a g e www.ijacsa.thesai.org issues. Compared to other algorithms, the exploration and exploitation ability of JO algorithm is very impressive and superior. And it has a very good balance among exploitation and exploration.

D. Recommendation Analysis
The following section gives the recommendation analysis in terms of a graphical user interface (GUI). Similarity-based outcomes are taken here for the statistical measures and compared with the existing classifiers. The proposed model DAE-SR achieved better performance, and it produced accurate recommendation outcomes. Recommendation of the good restaurants name, location and the reviews of the recommender is displayed in this analysis. Fig. 11 shows the content based restaurant recommendation. The recommendation text box shows the restaurant name, location, day and time also the customer reviews. There are large amount of data is there. The user's information that is related to the restaurant is shown by GUI.

E. Discussion
The proposed study introduces a DL based restaurant recommendation system named DAE-SR which extracts user preferences from online comments and recommends top restaurants to the users. With the technological advances and changes in human lifestyle makes the selection of appropriate restaurants as very difficult due to the wide range of ethnicities, ingredients, culinary styles, personal tastes and cultures. The proposed DAE-SAR based recommendation systems offers valuable decisions to users on where to eat. The proposed DAE-SR model attained an increased accuracy of 98.33% whereas the existing models RNN, LSTM, Ensemble, GRU, BiLSTM and RS-LOD gained lower results such as 86%, 89%, 92.36%, 90%, 89% and 84.1% respectively. Thus, the presented model obtained improved accuracy due to better convergence, low run-time and GUI modelling.

V. CONCLUSION
Recommender system provides precise online recommendations to the web users, it can also be known as knowledge management and information distribution tool. In this paper, a novel distributed framework for context aware recommender system is proposed based on text mining and machine learning algorithms. Based on the user opinion feedback reviews from the social networking services the top restaurants are recommended to the customers. The existing works on recommendation system proved that the use of individual CF based recommendation provide recommendations only via ratings/user interests and items. In this work, a new RS model is presented which proved that the contextual information is more relevant in the generation of recommendations under diverse conditions such as ubiquity and mobility. The proposed DAE-SR classifier offers suitable suggestions to users on the restaurant recommendations compared to the existing DNN, KNN and ANN methods. Experimental outcomes reveal that the proposed DAE-SR based recommendation system outperforms the other classifiers in terms of accuracy, precision, and recall values based on the similarity measures. The accuracy (98.33%) of proposed strategy DAE-SR outperforms the other strategies with cosine similarity. The main advantage of DAE-SR is, it may not suffer from cold-start issue and work well with sparse datasets. The limitation of this study is that only one dataset is used for analysis. Moreover, it is limited by the total amount of available content for providing meaningful recommendation. In future work, we can evaluate our proposed method with different datasets for different applications.