Research On Collaborative Filtering Based On Graph Neural Networks

In the information explosion era,the recommendation system can filter redundant information for users by mining their interests,and help them retrieve the goods and movies that they' re interested in.Collaborative filtering is one of the most influential recommendation algorithms based on the similar relationship between users and goods.How to further improve the recommendation quality on the idea of collaborative filtering is a hot issue.In recent years,the rapid development of collaborative filtering algorithm based on graph convolutional neural network provides a new idea for recommendation algorithm.This kind of algorithm reconstructs the interaction history between users and goods into graph structure data,and then uses graph convolutional structure to mine the deep relationship between users and goods,so as to improve the recommendation performance of collaborative filtering algorithm.Based on the above ideas,we proposes a new collaborative filtering model based on graph convolutional neural networks,which has two innovations.Firstly,based on the attention mechanism,we designs a new graph convolutional layer and simplifies it according to the characteristics of the recommended scene,so as to adaptively extract the high-order relationship between users and goods;Secondly,aiming at the over-smoothing problem,we proposes a new graph data enhancement method,which randomly shields the information transmission between nodes according to the graph data density,so as to enhance the robustness of recommendation model.In this thesis,comparative experiments with the state-of-the-art methods is carried out on three data sets from the real scene.Our experiments show that the simplified graph attention mechanism and the data enhancement method considering the density of graph can significantly improve the recommendation performance of our model.Compared with the baseline model,the proposed model has significantly improved the recall and the NDCG.