Information Compensation Graph Contrastive Learning for Recommendation

Abstract Applying graph convolutional neural networks to collaborative filtering is a novel approach pertaining to recommendation systems currently, which has afforded suitable results. However, certain problems still limit the performance of graph collaborative filtering, such as the data uniformity problem. In other words, the quality of embedding the expression of different data after multiple convolutions is reduced, leading to the decline of push model performance. In this paper, we propose self-supervised contrastive learning using global information compensation of feature embeddings, which can effectively alleviate the problem of data uniformity and improve model robustness. Simultaneously, we also propose a graph convolution method using local cooperative propagation to improve the performance of the recommendation model. This embedding calculation method for local cooperative propagation can maximize the influence of low-layer embedding on high-layer embedding, thereby improving the high-layer embedding uniformity. Experiments show that compared with the baseline, our model exhibits significantly improved performance on the three public datasets. Partially on the ML-1M dataset, the proposed ICCL exhibits a performance improvement of 7.96%, proving that our method is valid and explainable.