Research Of Personalized Recommendation Based On Social Review Data

With the development of the Internet,information overload has been an emergency issue,which has lead to the wide usage of personalized recommendation algorithms among many online platforms.Among all personalized recommendation algorithms,Collaborative Filtering(CF)methods have achieved great attention because of their effective performance,especially the latent factor-based models.The key idea of such latent factor-based models is mapping users and items into the latent embedding space,and the predicted rating of the target user-item pair can be calculated with their learned latent embedding vectors.However,because of the data sparsity issue and the lack of semantic information for users and items,the performance of such models is limited.Moreever,the explainability of recommendation models is still under-development.Therefore,in this thesis,we will focus on the study of improving the accuracy and explainability of the personalized recommendation model.Along with the development of the Internet,the Social Network Services(SNS)platforms have also been significantly developed.Many users share their daily activities on these platforms,such as their consumption records,reviews of restaurants they visited,and their social relations with others,which is helpful for boosting the accuracy and explainability of the personalized recommendation model.Among these different kinds of data,we mainly focus on social network data and review data because of the following reasons.First,the social network data and review data are easier to be collected.Second,these user-related data reveals users’ interests and behavior preferences from different perspectives.Third,these two kinds of data are public available,which is helpful of avoiding the privacy issue.However,how to better utilize such data is still not wellexplored.For example,for the social recommendation tasks which aim to improve the accuracy of recommendation models based on the social relations data,most researchers design simple and shallow linear models to model the influence of social neighbors.However,in the real world,the influence of social neighbors is very complex,and shallow and simple models may fail to model complex social relations.For the review-based recommendation model,most researchers argue that the main challenge is how to eliminate the semantic gap between the review text and the learned review representation.However,apart from the semantic gap,there also exists a preference gap between the review text and users’ preference representations.How to eliminate the preference gap is still unexplored.In this thesis,we aim to study how to utilize the social network and review data to enhance the accuracy and explainability of personalized recommendation models.To address the limitations of current works,i.e.,the complex social influence among social users can not be captured by the shallow and simple network of conventional social recommendation models,and the study of mitigating the preference gap between review data and users’ preferences is not well-developed,our proposed work contains following three parts: 1)Modeling users’ interest preference based on social influence(modeling the complex social relationships and improving the accuracy of recommendation model);2)Modeling the dual relationship between review text and rating data(mitigating the preference gap and improving the accuracy and explainability of recommendation models);3)Explainable recommendation based on unsupervised item aspects extraction(mitigating the preference gap and improving the accuracy and explainability of recommendation models).Following are the details of our proposed works.1.Modeling users’ preference based on social influence: we aim to study how to model the complex social influence and how to boost the accuracy of recommendation models based on users’ social relations.We assume that users’ preferences would change because of their social neighbors’ influence.Its obviously that users’ preference change over time until reaching steady.For the users whose preference are steady,their preferences are not only influenced by their first-order social neighbors,but also their higher-order social neighbors.The process that social users’ preference change over time can be treated as information propagation process in social networks.Luckily,the graph convolutional network is good at modeling the information propagation process in graph structure.In this thesis,we also utilize the graph convolutional network to model the evolution of social users’ preferences.Comparing with the convenient shallow and simple social recommendation models,the extensive experiments on two real-world datasets show the effectiveness of our proposed model in personalized recommendation task.2.Modeling the dual relationship between review text and rating data: we aim to mitigate the preference gap between review text and users’ preference,and boost the accuracy and explainability of the personalized recommendation models.The reviews describe the users’ feelings to different aspects of the items.For review-based recommendation task,reviews are used to enrich the preference representations of the users.More accurate users’ representations would improve the accuracy of the rating prediction model.Actually,if the recommendation model performed better in the rating prediction task,the preference gap between the review representations and the target user’s preference representations could be better mitigated.Generally,the ratings are treated as the ground truth of the rating prediction model.Because ratings data is very sparse,the performance of the rating prediction model is limited.To address such issue,we incorporate more supervision signal to constrain the optimization of rating prediction model.Users usually rate items by rating and writing reviews.And two dual tasks can be built on such two kinds of user behaviors,rating prediction task can be treated as the primay task,and review generation task can be treated as the dual task.Inspired by the dual learning theory,we can find there exists probabilistic duality property between the models of these two tasks.In the training stage,converting the probabilistic duality property into a regularization term,can lead the learned two models to better ones.The extensive experiments on two real-world dataset also show the effectiveness of the two models in rating prediction task and review generation task,which also shows the preference gap between the reviews and users’ preference can be better mitigated.3.Explainable recommendation based on unsupervised item aspects extraction: like the previous one,we aim to mitigate the preference gap between the users’ reviews and their preference representations.And the difference between this part and the previous part is,in this part,we focus more on the study of training a text generation model to generate more fine-grained explanation text.As the reviews contain many fine-graind information,such as users’ feelings to items’ different aspects,a natural idea is extracting the item aspects from the reviews.Then,a fine-grained explainable text generation model can be designed based on the extracted item aspects.However,it is not trivial,as the original reviews contain many endorsement text,which are not helpful to explain the users’ behavior intention,and the item aspects data is not available.To address such issues,we utilize an unsupervised aspect extraction module to extract item aspect information from the review data.Then,the users and items can be mapped to the aspect space with their related review data.Based on the users’ and items’ representations in the aspect space,we design a rating prediction module and explainable recommendation module.And the above modules are optimized in a unified framework.In the optimization stage,any two modules can enhance each other,and the performance of all modules can also be improved.Extensive experimental results on three real-world datasets demonstrate the superiority of our proposed model for both recommendation accuracy and explainability,which alsoshows the preference gap between the reviews and users’ preference can be better mitigated.In summary,to address the data sparsity issue and lack of explainability of conventient recommender models,we study how to boost the accuracy and explainability of the personalized recommendation models based on the users’ social relations and reviews text.For the conventional social recommendation models which utilize users’ social relations,the researchers usually design shallow and simple networks to model the social users’ influence,while in the real world,the social users’ influence is complex,and the conventient social recommendation models may fail to work.And for the conventional review-based recommendation models which utilize users’ reviews text,the researchers focus on the study of mitigate the semantic gap between the reviews text and their representations,while ignore the preference gap between the reviews text and the users’ preference representations.To address the limitations of conventional models,we study how to boost the performance of the personalized models based on the development of users’ preference modeling and behavior intention modeling in data mining area,social network analysis,text analysis,Graph Neural Network(GNN),dual learning,multi-task learning techniques,and so on.Extensive experiments on multiple real-world datasets show the effectiveness of our proposed models for recommendation accuracy and explainability.