Contrastive Learning based on Diffusion Graph for Implicit Collaborative Filtering

Recently, collaborative filtering based on contrastive learning has been intensively researched in the recommendation community, where positives are user-item interactions and negatives are selected from uninteractions. Although most studies have focused on negative sampling strategy and loss function design, few have investigated how positives would impact recommendation performance. Indeed, it could happen that some user-item interactions are kinds of misclicks, acting as noises during contrastive learning. This paper studies how to mitigate the effects of misclicks via the feature mixup technique and improves all components of contrastive learning. Taking the user-item interaction matrix as a bipartite graph, our proposed CLDG4CF model first obtains the associate matrix of its diffusion graph via the personalized random walk with restart. The associate matrix to some extent describes the potential interaction likelihoods between users and items in the convergence state, which is used for positive generation, negative sampling, and weighted loss in contrastive learning. In particular, we randomly select interacted items and weighted mixup their encodings as a generated positive. We select uninteracted items as negatives based on the item-item part in the associate matrix according to their association coefficients to those already selected interactions. Finally, we use their association coefficients to compost a weighted contrastive loss in contrastive learning. Experiments on public datasets demonstrate the effectiveness of our design in terms of the improved recommendation performance over the peer contrastive learning approaches.