Are Graph Augmentations Necessary? Simple Graph Contrastive Learning for Recommendation (2112.08679v4)

Abstract: Contrastive learning (CL) recently has spurred a fruitful line of research in the field of recommendation, since its ability to extract self-supervised signals from the raw data is well-aligned with recommender systems' needs for tackling the data sparsity issue. A typical pipeline of CL-based recommendation models is first augmenting the user-item bipartite graph with structure perturbations, and then maximizing the node representation consistency between different graph augmentations. Although this paradigm turns out to be effective, what underlies the performance gains is still a mystery. In this paper, we first experimentally disclose that, in CL-based recommendation models, CL operates by learning more evenly distributed user/item representations that can implicitly mitigate the popularity bias. Meanwhile, we reveal that the graph augmentations, which were considered necessary, just play a trivial role. Based on this finding, we propose a simple CL method which discards the graph augmentations and instead adds uniform noises to the embedding space for creating contrastive views. A comprehensive experimental study on three benchmark datasets demonstrates that, though it appears strikingly simple, the proposed method can smoothly adjust the uniformity of learned representations and has distinct advantages over its graph augmentation-based counterparts in terms of recommendation accuracy and training efficiency. The code is released at https://github.com/Coder-Yu/QRec.

Simple Graph Contrastive Learning for Recommendation

The paper "Are Graph Augmentations Necessary? Simple Graph Contrastive Learning for Recommendation" questions the necessity of graph augmentations in contrastive learning (CL) within recommender systems. It introduces a simplified method that forgoes traditional graph augmentations to achieve enhanced recommendation performance.

Overview

Graph neural networks (GNNs) have become fundamental in processing recommendation data. Typically, CL applies to recommendations by augmenting user-item graphs with structural perturbations like node or edge dropout, to learn invariant representations. The paper investigates whether these augmentations are essential or if CL alone can provide sufficient representational power.

Key Findings

1. Impact of Graph Augmentations: The paper conducts experiments revealing that the real driver of performance improvement in CL-based recommendation models is the CL loss rather than graph augmentations. It shows that models without graph augmentations (SGL-WA) can outperform variants that use them.
2. Uniformity and Popularity Bias: It identifies that optimizing CL loss results in a more evenly distributed representation space, thereby implicitly mitigating popularity bias. This effect helps improve the generalization ability of recommendations by reducing emphasis on popular items.
3. Proposed Method (SimGCL): The authors propose SimGCL, which enhances CL by introducing uniform noise to the embedding space instead of complex graph augmentations. This method regulates the uniformity of the learned representations more efficiently and effectively.

Experimental Evaluations

• Datasets: Experiments were conducted on three datasets: Douban-Book, Yelp2018, and Amazon-Book.
• Results: SimGCL demonstrated superior performance compared to its augmented counterparts and had distinct advantages in terms of recommendation accuracy and training efficiency.
• Model Analysis: The paper includes an analysis of convergence speed and running time, illustrating SimGCL's efficiency. The method significantly reduces the training epoch time compared to other CL-based techniques.

Implications

The research has significant implications for developing efficient recommender systems. By eliminating the need for graph augmentations, SimGCL simplifies the architecture and reduces computational overhead. This approach not only improves scalability but also enhances the system's ability to provide unbiased recommendations.

Future Directions

The paper opens avenues for further experimentation with different noise models and their impact on representation uniformity. Future work could explore dynamic adjustment of noise levels throughout training or leverage other self-supervised learning techniques to refine user-item embeddings.

Conclusion

This work effectively challenges the traditional reliance on graph augmentations in CL for recommendations, introducing an innovative and efficient alternative. The findings and proposed methodology offer a promising direction for advancing the design of recommender systems, providing a balance between complexity and performance.