Recommendation Algorithm Based on Multi-view Fusion Cross-layer Contrastive Learning
Existing recommendation models based on graph comparison learning usually use only one view enhancement method in graph data enhancement,ignoring the limitations of a single method.In contrastive learning,only a pair of views from the same node are usually compared,which do not fully utilize the different layer embeddings of each view.To this end,this study proposes a recommendation algorithm framework based on Multi-view Fusion Cross-layer Contrastive Learning(MFCCL).MFCCL constructs two global views using random edge drop and random noise addition enhancement methods,local views using Singular Value Decomposition(SVD),and three global and local views using three different view enhancement methods to achieve effective user representation.Simultaneously,a new MFCCL method is proposed,which embeds different layers of two global views through parallel and cross fusion methods for comparison,to obtain more feature information.Combining MFCCL with global-local view contrastive learning aims to jointly optimize the model and improve recommendation performance.Experiments are conducted on three publicly available datasets:Yelp,Tmall,and Amazon-book,and the results demonstrate that MFCCL is effective and feasible in recommendation tasks.Compared to the baseline model SimGCL,which had the best neutral performance in the comparative model,MFCCL exhibited better performance in the three datasets.The Recall@20 gains reached 15.0%,13.3%,and 28.7%,and the NDCG@20 values increased by 14.3%,13.2%,and 29.6%,respectively.
国家科技期刊平台
NSTL
万方数据
维普
