Recommendation Based On Graph Convolutional Network And Knowledge Graph

In recent years,the Internet has developed rapidly and explosive growth of information,recommendation system is playing an important role in helping users to find interesting content and helping enterprises to promote their products.Moreover,recommendation system has been widely applied in many fields such as finance,consulting,social contact,film,education and so on.In practical applications,most recommender systems adopt collaborative filtering algorithm,input user item interaction data,and capture the collaborative signal of user-item interaction through interaction function.Although this approach can cope with most of the recommended tasks,there are some problems: 1)Interaction functions,graph convolutional network is currently a hot research direction,but there have been smoothing problems;2)Data sparsity and cold start caused by limitations of input data of collaborative filtering algorithm.In order to solve the above problems,this thesis improves the traditional model through two methods: 1)The improved graph convolutional network is a multi-subgraph convolutional network and the random drop edge mechanism is introduced to alleviate the over-smoothing and over-fitting problems.The main idea is to analyze it from the perspective of recommendation domain and divide the subgraph according to the feature of users,so that the high-order embedding propagation is only among users with similar feature,and avoid the noise in high-order embedding propagation;2)The text convolution layer is used to effectively connect the entity in the knowledge graph and the word embedding as the embedding of the movie.The convolutional network is mainly used to regard words and entities as multiple channels and ensure their alignment in the process of convolution.In this method,the knowledge level connection between movies is integrated into the improved multi-subgraph convolutional network,and the reasonable expansion of the multi-subgraph convolutional network is realized,so as to solve the data sparse and cold start problems.In this thesis,two improved methods are experimented respectively.The research results showed that: 1)On the real datasets of four different fields,the evaluation indexes were respective improved by 2.5%,4.5%,13.1% and 4.08%,compared with the traditional recommendation model.2)Compared with the recommendation model based on graph convolution,the improved multi-subgraph convolutional network recommendation model can carry out high-order embedding propagation in the deeper graph convolutional network and obtain better recommendation effect.3)Experiments are conducted on two public datasets in the movie field.The results show that the performance of the text convolution layer adopted in this paper,which integrates entities and words into the improved multi-subgraph recommendation model,is significantly better than that of the traditional recommendation model.In addition,from the perspective of knowledge graph,this thesis provides possible explanations for the recommendation results.Finally,this paper builds a personalized movie recommendation system based on Django framework.The system uses the above improved model to realize personalized recommendation for the movie field.