DimCL: Dimension-Aware Augmentation in Contrastive Learning for Recommendation

Contrastive learning (CL) has achieved remarkable success in addressing data sparsity issues in collaborative filtering (CF) for recommender systems (RSs). The key principle is to generate different augmented views given a user-item interaction graph. However, prior endeavors mainly focus on performing augmentation via stochastic functions, e.g., by injecting perturbations into different hidden dimensions uniformly. Without fine control, the hidden representations of augmentations may contain noisy dimensions that are harmful to CL and irrelevant to RSs. Removing dimension-specific noise is a challenging task due to the following two major bottlenecks. It is difficult to (i) distinguish different dimensions' efficacy for CL and (ii) bridge the semantic gap between CL and RSs. Overlooking these limitations may cause redundant, false-positive, and irrelevant noise in hidden dimensions of the augmented views. In this paper, we solve the above challenges from the perspective of robust learning and curriculum learning, and propose a novel Dimension-aware augmentation in Ceontrastive Leearning for recommendation (DimCL). In DimCL, we first theoretically analyze the easiness and hardness of different dimensions for CL and RSs. With thorough analysis, we propose two propositions, which reveal that the gradients of different dimensions of augmented views are potentially related to the learning difficulty of optimizing CL and RSs. The comparison of gradients can provide detectable signals to reflect the efficacy of different dimensions for CL and the semantic gap between CL and RSs. Based on the analysis results, we devise three denoising factors, which can help DimCL to identify hard-to-learn dimensions as redundant or false-positive noise and pinpoint dimensions in different augmented views with inconsistent difficulties of RSs as irrelevant noise without requiring additional supervised labels. After denoising, DimCL can remove dimension-level noise to reduce unnecessary difficulty, making CL easier and maintaining more consistent difficulty in RSs. Extensive experiments on four public datasets demonstrate DimCL's superiority and flexible applications over various traditional and CL-based CF methods. The source code is publicly available online at https://github.com/zc-97/DimCL.