Supervised contrastive learning for recommendation

In the recommendation system, collaborative filtering methods based on graph convolution network can explicitly model interaction between nodes of user–item bipartite and effectively use higher-order neighbor information. However, its representations are very susceptible to noise interaction. response this problem, SGL explored self-supervised learning improve robustness GCN. Nevertheless, contrastive framework it applied does not consider specificity task uncertainty fully. order solve above problems, we propose a paradigm called supervised (SCL) convolutional neural network. We carefully design SCL guided by basic idea that users with similar histories have interests preferences. Specifically, will calculate similarity different user side item respectively during data preprocessing firstly. And then when applying learning, only augmented samples be regarded as positive samples, but also certain number which is treats other in batch negative samples. purposefully makes learned close each feature space. addition, address node interaction, new augment method replication. apply most advanced LightGCN. Empirical research ablation study Gowalla, Yelp2018, Amazon-Book datasets prove effectiveness, accuracy,