Supervised Contrastive Learning for Recommendation
share
Compared with the traditional collaborative filtering methods, the graph convolution network can explicitly model the interaction between the nodes of the user-item bipartite graph and effectively use higher-order neighbors, which enables the graph neural network to obtain more effective embeddings for recommendation, such as NGCF And LightGCN. However, its representations is 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. Although effective, we found that SGL directly applies SimCLR's comparative learning framework. This framework may not be directly applicable to the scenario of the recommendation system, and does not fully consider the uncertainty of user-item interaction.In this work, we aim to consider the application of contrastive learning in the scenario of the recommendation system adequately, making it more suitable for recommendation task. We propose a supervised contrastive learning framework to pre-train the user-item bipartite graph, and then fine-tune the graph convolutional neural network. Specifically, we will compare the similarity between users and items during data preprocessing, and then when applying contrastive learning, not only will the augmented views be regarded as the positive samples, but also a certain number of similar samples will be regarded as the positive samples, which is different from SimCLR who treats other samples in a batch as negative samples. We term this learning method as Supervised Contrastive Learning(SCL) and apply it on the most advanced LightGCN. In addition, in order to consider the uncertainty of node interaction, we also propose a new data augment method called node replication.
READ FULL TEXT