Research On Recommendation Method Based On Deep Learning By Integrating Review Information

With the advent of the Internet era,information overload has become a severe challenge for human society.In this context,the recommender system has become an indispensable part of online service platforms such as e-commerce websites,social media and news portals.The main goal of recommender system is to solve the problem of information overload by mining users’ interests.At present,mainstream recommender systems usually use historical interaction data between users and items to make recommendations.Although these recommender system based on interactive data performs well,its performance will decline sharply when interactive data is missing or new items are encountered.For these problems,the recommendation method integrating user reviews has become popular in recent years.User reviews are an important online consumer behavior data,which can express users’ experience and emotional categories of items.The recommendation method integrating user reviews tries to use user reviews as an auxiliary data source to solve the problems of sparse user interaction data and cold start faced by traditional recommendation systems.At present,there are still some shortcomings in the recommendation method integrating user reviews:(1)Most of the existing recommendation method integrating user reviews do not consider the special importance of user reviews on the target item(RT)in building user preferences;(2)The existing methods mostly adopt the same data format and processing method,ignoring the essential difference between the user and the item reviews;(3)Most existing methods ignore the time stamp and collaborative features of reviews.In view of the above shortcomings,this dissertation will study from the following three aspects:(1)MAN: Main-auxiliary network with attentive interactions for review-based recommendationFor the initial stage of the e-commerce platform with a small number of people,it is necessary to pay special attention to the importance of RT in building user preferences,so as to capture users’ liking and intuitive feelings about the target item,and thus improve the accuracy of recommendation results.Therefore,this dissertation propose a Main-Auxiliary Network(MAN)based on deep learning for item recommendation.Specifically,MAN uses the auxiliary network to focus on the purification of RT at the word level and assists the main network in generating the predicted value of RT.The main network deals with the user-item interaction according to the relationship between the user multiaspect features and the item as the most prominent aspect feature and then generates the final rating prediction.Extensive experiments on five public datasets show that MAN outperforms the state-of-the-art methods.(2)Asymmetric multilevel interactive attention network integrating reviews for item RecommendationFor the medium stage of the e-commerce platform with the gradual increase of the number of people,it is necessary to adopt different data formats and processing methods for the review sets.Because the user reviews are made by the user on many items they have purchased,it reflects the broad interests of the user and belongs to the heterogeneous data type.The item reviews are made by different users according to the attributes of the item,which belongs to the homogeneous data type.It is conducive to building more accurate the user and the item features and thus producing better recommendation performance.Therefore,this dissertation propose a novel Asymmetric Multi-Level Interactive Attention Network(AMLIAN)integrating reviews for item recommendation.AMLIAN can predict precise ratings to help the user make better and faster decisions.Specifically,to address the essential difference between the user and the item reviews,AMLIAN uses the asymmetric network to construct user and item features using different data forms(document-level and review-level).To learn more personalized user-item interaction,the user ID and item ID and some processed features of user reviews and item reviews are respectively used for multilevel relationships.Experiments on five real-world datasets show that AMLIAN significantly outperforms state-of-the-art methods.(3)A graph attention network with contrastive learning for temporal review-based recommendationsFor the later stage of the e-commerce platform with a large number of people,the timestamp and collaborative features of reviews are very important.Paying attention to the former can capture the dynamic changes of user preferences and item attribute,while paying attention to the latter can enhance the accuracy and novelty of recommendation.To overcome these problems,this dissertation propose a novel Graph Attention Network with Contrastive Learning(GANCL)for temporary review-based recommendation.Specifically,to capture dynamic user preferences and high-order collaborative features,we design a user-item bipartite graph with time-series review information and ratings as its edges,and we then use the graph attention and different gating mechanisms to extract the corresponding features.To make full use of the limited interaction between the users and items,we use the comparative learning paradigm for the nodes and edges in the bipartite graph to more effectively model the user-item interaction.Extensive experiments on five public datasets show that the performance of GANCL significantly outperforms the state-of-the-art methods.