Research On Graph Attention Collaborative Filtering Algorithm Based On Comparative Learning

With the development of deep learning,collaborative filtering based on deep learning has become one of the mainstream algorithms for recommendation systems,among which collaborative filtering algorithms based on graph learning are the research hotspots in this field.Existing collaborative filtering methods based on graph learning integrate the bipartite graph structure of user-item interactions in the embedding process,but this method suffers from the problem of oversmoothing;meanwhile the method still does not alleviate the data sparsity and long-tail problem,and some cold items and user interactions are not fully learned in this case.In response to the above problems,the main work of this paper is as follows:(1)To address the problem of oversmoothing in traditional graph convolutionbased collaborative filtering models,this thesis proposes a graph convolutional coll aborative filtering model GACF based on a self-attentive mechanism.incorporating a graph-attentive mechanism to calculate neighbor weights in the message propagation process and fusing the interaction coding between users and items by weighting the attention coefficients,which can improve the performance while increasing the number of layers and avoiding the occurrence of oversmoothing.The experim ental results show that the method was able to improve its NDCG@20 and Recall@20 by 1.943% and 2.129% on each data set,respectively.(2)To address the long-tail problem and data sparsity problem of recommendation systems,this paper uses self-supervised learning to preprocess the main recommendation task based on the GACF model to discover the self-supervised signals on the recommendation data,and proposes the composite contrast learning frame work CAGL.first,the edge prediction function of a graph is used to predict the corresponding edge probabilities,and the operation of discarding and adding edges is implemented deterministically according to the probability matrix Data augmentation is then performed,and then multi-task learning is performed to calculate the common supervised learning loss and contrast learning loss respectively to jointly optimize the recommendation model.Experimental results are shown in the larger dataset Yelp 2018 and Alibaba-i Fashion,the NDCG@20 was increased by 2.79%and 2.41%,respectively.Figure [22] Table [6] Reference [71]...