Research On Recommendation Algorithm Based On Graph Convolutional Network And Fine-grained Sentiment Analysis

With the rapid development of the Internet,people can get more and more information from the Internet.The surge of information will also bring about the problem of information overload.Although people can find the information they need more easily,the useless information for them also surges.In order to solve the problem of users' information overload,recommendation system has emerged.In recommendation system technology,the collaborative filtering algorithm is the most widely used algorithm.However,the development of collaborative filtering algorithm still faces many issues such as data sparsity and cold-start.In this dissertation,aiming at the defects of collaborative filtering algorithm and the shortcomings of existing methods,an improved scheme is proposed to achieve more accurate recommendation.The core idea of the collaborative filtering algorithm is to calculate the similarity between users based on the interaction between the user and the item,and then predict other items that the user may be interested in based on this.In some existing studies,graph neural networks are introduced into the model to measure the similarity between users,and good results have been achieved.However,most of the existing work did not consider the importance of different neighborhoods,and failed to distinguish between neighborhoods.At the same time,the user's review text contains more emotional information.Many researchers are also trying to use the emotional information contained in the review text for recommendation tasks.Existing work fails to consider the integration of many features in the review text.Measure user interest.Therefore,this dissertation focuses on the data sparseness problem faced by collaborative filtering algorithms,and conducts research from two aspects: neighborhood distinction and multi-feature fusion.This dissertation mainly does the following work:1.Research on recommendation model combining lightweight graph convolutional network and attention mechanism.The commonly used collaborative filtering algorithms are either based on matrix factorization algorithms to mine the linear relationship between users and items for similarity measurement,or use neural networks to build non-linear functions.Most of these algorithms ignore the high-level connection relationship between users and items,and the high-level connection relationship can obtain more collaborative filtering signals.Graph convolutional networks can effectively capture these high-order connections.However,the existing model uniformly gives a fixed value to the weights of all neighborhoods,and fails to distinguish the domains according to their importance.This dissertation solves this problem by introducing an attention mechanism,and proposes a recommendation algorithm that combines a lightweight graph convolutional network and an attention mechanism.Finally,experiments were conducted on three real data sets of Gowalla,Yelp2018 and Amazon-book to verify the feasibility of the model.2.Research on recommendation algorithms incorporating fine-grained sentiment analysis.In order to alleviate the problem of data sparsity,the use of user comment text information is also a popular research method.For the many features of the review texts,if they can be effectively used and integrated,the performance of the recommendation system will be greatly promoted.This dissertation mainly selects the following three characteristics to be integrated:the relative sentiment scores for different evaluations;the extent to which the user review text helps other users,that is,the influence of the review;whether the user's score is consistent with the sentiment expressed by the review,that is,consistency.At the same time,it is recommended by combining the interest measurement method at the coarse-grained level.Finally,experiments were conducted on the hotel review data set.The experimental results show that in terms of data sparsity,the integration of many features of the review text can effectively cope with the problem,and the recommendation effect of the recommendation algorithm has also been further improved.