Improving Graph Collaborative Filtering with Neighborhood-enriched Contrastive Learning

Abstract

Recently, graph collaborative filtering methods have been proposed as an\neffective recommendation approach, which can capture users' preference over\nitems by modeling the user-item interaction graphs. In order to reduce the\ninfluence of data sparsity, contrastive learning is adopted in graph\ncollaborative filtering for enhancing the performance. However, these methods\ntypically construct the contrastive pairs by random sampling, which neglect the\nneighboring relations among users (or items) and fail to fully exploit the\npotential of contrastive learning for recommendation. To tackle the above\nissue, we propose a novel contrastive learning approach, named\nNeighborhood-enriched Contrastive Learning, named NCL, which explicitly\nincorporates the potential neighbors into contrastive pairs. Specifically, we\nintroduce the neighbors of a user (or an item) from graph structure and\nsemantic space respectively. For the structural neighbors on the interaction\ngraph, we develop a novel structure-contrastive objective that regards users\n(or items) and their structural neighbors as positive contrastive pairs. In\nimplementation, the representations of users (or items) and neighbors\ncorrespond to the outputs of different GNN layers. Furthermore, to excavate the\npotential neighbor relation in semantic space, we assume that users with\nsimilar representations are within the semantic neighborhood, and incorporate\nthese semantic neighbors into the prototype-contrastive objective. The proposed\nNCL can be optimized with EM algorithm and generalized to apply to graph\ncollaborative filtering methods. Extensive experiments on five public datasets\ndemonstrate the effectiveness of the proposed NCL, notably with 26% and 17%\nperformance gain over a competitive graph collaborative filtering base model on\nthe Yelp and Amazon-book datasets respectively. Our code is available at:\nhttps://github.com/RUCAIBox/NCL.\n