Supervised contrastive learning for recommendation

Abstract

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