Dense Connection Graph Contrastive Learning for Recommendation

Abstract The application of a graph convolutional network (GCN) to collaborative filtering (CF) is a new direction of recommendation system and has achieved good results. However, the problem of data uniformity, that is, the quality of embedded expression of different data after multiple convolutions, still persists. In this paper, we propose a convolution method using dense connection, which can effectively reduce data uniformity and improve the performance of the recommended model. This dense connection embedding calculation method can maximize the influence of low-order embedding on high-order embedding, thereby improving the uniformity of higher-order embedding. At the same time, the noise problem also affects the quality of embedded expression. We introduce contrastive learning into graph CF to alleviate the noise problem. Contrastive learning optimizes contrastive loss by reducing the distance between positive samples and increasing the distance between negative samples. We use a contrastive learning method through graph perturbation. Specifically, we randomly lose the edges of the graph twice to make contrastive learning between two graphs. At the same time, we also compare different nodes of each subgraph. This contrastive learning method improves the performance of the recommendation model. Experiments show that our model has significantly improved on multiple open datasets. Compared with the baseline, our model has 14% and 31% performance improvements on yelp2018 and book-crossing datasets, respectively, proving that our changes are effective and interpretable.