融合多尺度因果去偏与自监督对抗判别的图推荐模型OA

In graph recommendation systems,user interactions are frequently disturbed by confounding factors such as popu-larity trends and community effects,which distort authentic user preference features and degrade long-tail item recall perfor-mance.To tackle such challenges,this study proposed a graph recommendation model named MCAP that combined multi-scale causal debiasing and self-supervised adversarial purification.The model constructed a parallel dual-branch graph convolutional structure,independently extracted user interest embeddings from user-item interaction graph and bias embeddings from bias correlation graph,and separated the two representations via decoupling loss.On this basis,it established a multi-scale causal debiasing chain to conduct causal intervention on interest embeddings according to bias information at embedding level,and suppress bias amplification at scoring level with global topological features of bias association graph.It further introduced a self-supervised adversarial discrimination module to eliminate residual confounding interference,and adopted gradient reversal layer to realize adversarial training between interest branch and discriminator,improved the purity of unbiased preference rep-resentations.Extensive experiments on MovieLens-10M and Netflix datasets demonstrate that MCAP achieves superior recall@20/50 and NDCG@20/50 performance compared with mainstream debiasing recommendation baselines.The results verify that the proposed method effectively alleviates confounding interference and boosts overall recommendation quality.