Research On Personalized Recommendation And Interpretability By Integrating Semantics Of Reviews

In the era of informaiton explosion,recommender systems as information filtering systems have become important components of various Internet applications.Existing recommendation methods usually model users and items based on interactive behaviors such as ratings and clicks,but they often suffer from data sparsity problem,leading to insufficient representation learning of users and items.This would cause performance degradation in recommendation and a lack of interpretability.User-generated content(e.g.,reviews)can alleviate the problem of data sparsity;Mining review semantic information could obtain user preferences and item characteristics,and enhance the representation learning for users and items,which plays an important role in recommendation.This thesis focuses on personalized recommendation and interpretability integrated with reviews.From review usability,precision and interpretability of recommendation,this thesis has achieved the following contributions:First,a review helpfulness prediction model based on dual personalized attention is proposed.The personalization features of users and items are fused to enhance the semantic representation of reviews.A dual personalized attention is designed.The core is that the query vectors are derived from the ID embeddings of users and items,which is used to select more related and useful words in reviews.In this way,the personalized features of reviews could be learned.The method could effectively improve the performance of review helpfulness prediction and help the recommender system identify more valuable reviews.Second,a hybrid recommendation model that incorporates rating patterns and review information is proposed for rating prediction.The inherent complementarity between ratings and reviews of users and items is exploited: reviews can alleviate the sparsity of rating data,while rating patterns can explicitly portray the sentiment of users towards items.An attention module that fuses rating patterns into review learning is proposed,which could focus on more important reviews under the orientation of rating patterns.In this way,the proposed method could obtain more accurate representations for users/items and improve the precision of rating prediction in recommendation.Third,a rating prediction model for learning multi-grained representations of users/items is proposed,aiming to characterize the three crucial features including “diversity”,“contextawareness” and “personalization” for users and items from reviews.A novel three-tier attention model is designed to learn comprehensive features of users and items,including a review encoder based on word-level attention,an aspect encoder based on review-level attention,and a user/item encoder with aspect-level attention.The proposed method further improves the precision of rating prediction in recommendation.Fourth,a personalized review generation model for recommendation interpretability is proposed to provide interpretable reasons for recommendation.To solve the problem of scarcity of input information in review generation multiple sources of information such as attribute features and text features of users and items are fully utilized.Multiple encoders are designed to learn the writing style of users and features of items,which could provide rich context embedings for decoders.A decoder based on hierarchical multi-source gating mechanism is designed to select the context features that are relevant to the decoding state for review decoding.The proposed model could generate reviews with excellent performance in terms of fluency,readability,and personalization,which can provide interpretability for recommendation.Based on review semantic mining,the above research contents in this thesis,including review usability,precision and interpretability,enrich the research framework of review-based recommendation,and improve user experience of recommender systems.