Research On Graph Collaborative Filtering Enhanced Using Contrastive Learning

With the development of information technology,the massive amount of information on the internet has reached a saturated state.When users search for things they are interested in,they often become anxious and passive due to retrieving too much irrelevant information.The emergence of recommendation systems has alleviated this so-called"information overload" problem.Recommendation systems can proactively provide users with items they may be interested in based on their characteristics.Among them,graph collaborative filtering(referred to as graph CF)based on graph neural networks is a representative technology of recommendation systems.Collaborative filtering calculates the feature vectors of user items based on the interaction matrix constructed from the interaction records of user items,and calculates the user’s interest in items according to the feature vectors,thus achieving the recommendation effect.The graph convolution and message passing mechanism of graph neural networks can extract more structural information from the bipartite graphs formed by the interaction matrix,thereby helping collaborative filtering generate higher-quality node features.However,due to the sparsity of data sets in recommendation systems,graph neural networks cannot effectively extract information,and the model is challenged by data sparsity.In recent years,contrastive learning has developed rapidly.Its idea is to obtain expanded nodes through data enhancement,and then pull the feature representations of similar nodes closer while pushing away the feature representations of different nodes,thereby generating richer node features based on the original data set.Therefore,contrastive learning can alleviate the problem of data sparsity in graph CF to some extent.This paper will study how contrastive learning enhances graph neural network collaborative filtering.Specifically,the research contents of this paper are:(1)For the user-item bipartite graph,a large amount of information is often carried on the paths.Therefore,we propose a model named Node and Meta-Path Contrastive Learning for Recommender Systems(NPCRS).First,the model extracts meta-paths from the user-item bipartite graph and generates user and item meta-path views,which carry a large amount of path information.Then,we use contrastive learning to compare the meta-path views with the nodes on the original view,enabling the model to learn the information carried on the paths.We conducted repeated experiments on real datasets,and the results prove the effectiveness of the model.(2)In order to better extract information of similar nodes and enhance the generalization ability of the model,we propose a model named Contrastive Graph Collaborative Filtering with Graph Clustering and Perturbation(CGCF),which consists of a clustering information module and a perturbation module.The clustering information module classifies nodes by spectral clustering and k-means algorithms to obtain the feature vector of each type of node.We then use contrastive learning to pull similar nodes closer together to enable the model to learn more information about nodes of the same type.The perturbation module generates feature vectors with similar semantics by adding noise to the original feature vectors,and then enhances the robustness of the feature vectors through contrastive learning.We conducted experiments on two public datasets,verifying the effectiveness of the proposed model.(3)This study implemented a movie recommendation website based on the proposed model.Combining the explored algorithms,a recommendation module was designed that can calculate users’ interest characteristics based on their interaction history with movies.This generates personalized recommendation services for users.