Research On Deep Recommendation Algorithms Based On User Reviews

In modern commercial systems,the recommender system plays a vital role in helping users overcome the information overload problem.Traditional collaborative filtering methods are faced with a series of problems such as data sparsity problem when only using users' interactions to make recommendations.As an explanation of the user behavior,user reviews contain rich information.Using this information to help improve the accuracy of recommendation is of great significance in business.The existing review based recommendation method has some limitations in modeling review semantics and scalability.To overcome these problems,this paper uses deep learning methods to improve recommendation accuracy by researching a more effective way to organize the users' interactions and combining the review semantics with the characteristics of reviews considered.The work of the paper is as follows:(1)To model collaborative filtering signals more effectively,this paper proposes a deep collaborative filtering model based on joint auto-encoder.In this method,a bipartite graph rather than the rating matrix is used to represent users' historical interactions.With the graph structure considered,a model based on auto-encoders is used to model the collaborative signals more accurately to join learn the representation of users and items,which results in more accurate and robust recommendations.(2)Collaborative filtering methods,which only models the interaction itself without respecting to reasons behind it.To overcome the problem of collaborative filtering methods,the semantic characteristics of user reviews are firstly revealed.Then a review semantics based model for rating prediction is proposed,which links the semantic characteristics of user reviews with the making-decision process of users.This method draws on the ideas of generating adversarial networks and transfer learning to model the semantics of user reviews,which uses a single review rather than an aggregate review for training,thereby improving the accuracy of rating prediction and the scalability of the model.(3)Combining the ideas of the above two methods,a knowledge graph is used to organize users‘historical interactions and the semantics of user reviews,then a deep hybrid recommendation algorithm based on user reviews is proposed.This method first extracts the attribute aspect of the item mentioned by the user in the user review,then builds up a knowledge graph based on the interaction between the user and the item,and finally learns representations of the user and the item through the graph convolutional neural network to make a recommendation.Through a series of experiments with compared methods,the result shows: Firstly,it's more efficient to model the collaborative signals by regarding user-item interactions as a graph,which can enhance the performance of collaborative filtering methods in the data sparsity scenario.Secondly,It can improve the recommendation accuracy by considering the semantics of reviews and linking it with the making-decision process of users.Thirdly,it improves the model accuracy by formatting the review semantics and user-item interactions into a knowledge graph to learn the representations of users and items,which combines the strength of user-item interactions and user review semantics...