L^2CL: Embarrassingly Simple Layer-to-Layer Contrastive Learning for Graph Collaborative Filtering

TL;DR

This work tackles the limitations of graph contrastive learning in recommendation by eliminating reliance on data augmentations and deep architectures. It introduces L$^2$CL, an augmentation-free layer-to-layer contrastive learning framework that relies on a 1-hop, single-layer GNN to align embeddings across layers, reducing noise and improving efficiency. The authors provide information-theoretic guarantees that emphasize maximizing task-relevant information while limiting task-irrelevant noise, and validate the approach with extensive experiments on five real-world datasets, where L$^2$CL consistently outperforms state-of-the-art baselines and demonstrates robustness to data sparsity and faster training. The results suggest that simple, layer-based contrasts can effectively capture meaningful collaborative signals, offering a practical and scalable alternative to augmentation-heavy CL methods in graph-based recommender systems.

Abstract

Graph neural networks (GNNs) have recently emerged as an effective approach to model neighborhood signals in collaborative filtering. Towards this research line, graph contrastive learning (GCL) demonstrates robust capabilities to address the supervision label shortage issue through generating massive self-supervised signals. Despite its effectiveness, GCL for recommendation suffers seriously from two main challenges: i) GCL relies on graph augmentation to generate semantically different views for contrasting, which could potentially disrupt key information and introduce unwanted noise; ii) current works for GCL primarily focus on contrasting representations using sophisticated networks architecture (usually deep) to capture high-order interactions, which leads to increased computational complexity and suboptimal training efficiency. To this end, we propose L2CL, a principled Layer-to-Layer Contrastive Learning framework that contrasts representations from different layers. By aligning the semantic similarities between different layers, L2CL enables the learning of complex structural relationships and gets rid of the noise perturbation in stochastic data augmentation. Surprisingly, we find that L2CL, using only one-hop contrastive learning paradigm, is able to capture intrinsic semantic structures and improve the quality of node representation, leading to a simple yet effective architecture. We also provide theoretical guarantees for L2CL in minimizing task-irrelevant information. Extensive experiments on five real-world datasets demonstrate the superiority of our model over various state-of-the-art collaborative filtering methods. Our code is available at https://github.com/downeykking/L2CL.

Figures (9)

• Figure 1: Different layer-to-layer contrast.
• Figure 2: The left is an illustration of information in self-supervised learning. $L_1$, $L_2$, and $\mathcal{Y}$ represent the information of two layers and downstream tasks, respectively. The right is the asymmetric layer contrast strategy, which can alleviate the task-irrelevant noises in GCL (area C).
• Figure 3: The illustration of our proposed L$^2$CL, which jointly optimize one-hop contrastive learning loss and BPR loss.
• Figure 4: Performance on users of different sparsity degrees in terms of Recall@10. G1 denotes the group of users with the lowest average number of interactions.
• Figure 5: LightGCN