Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Symmetric Graph Contrastive Learning against Noisy Views for Recommendation

Chu Zhao, Enneng Yang, Yuliang Liang, Jianzhe Zhao, Guibing Guo, Xingwei Wang

TL;DR

This work tackles the problem that graph contrastive learning for recommendations can suffer when data augmentation creates noisy views that distort the original graph. It introduces Symmetric Graph Contrastive Learning (SGCL), a model-agnostic framework built on a symmetry-based loss that remains robust to noisy interference, and proves its compatibility with mutual information bounds via the Wasserstein Dependency Measure. The authors provide theoretical guarantees of noise tolerance and demonstrate that SGCL consistently improves recommendation accuracy on three real-world datasets, achieving up to 12.25% relative gains, while maintaining robustness under data sparsity and perturbations. These results offer a principled approach to robust self-supervised learning in graph-based recommender systems and suggest a practical, broadly applicable loss design for contrastive learning.

Abstract

Graph Contrastive Learning (GCL) leverages data augmentation techniques to produce contrasting views, enhancing the accuracy of recommendation systems through learning the consistency between contrastive views. However, existing augmentation methods, such as directly perturbing interaction graph (e.g., node/edge dropout), may interfere with the original connections and generate poor contrasting views, resulting in sub-optimal performance. In this paper, we define the views that share only a small amount of information with the original graph due to poor data augmentation as noisy views (i.e., the last 20% of the views with a cosine similarity value less than 0.1 to the original view). We demonstrate through detailed experiments that noisy views will significantly degrade recommendation performance. Further, we propose a model-agnostic Symmetric Graph Contrastive Learning (SGCL) method with theoretical guarantees to address this issue. Specifically, we introduce symmetry theory into graph contrastive learning, based on which we propose a symmetric form and contrast loss resistant to noisy interference. We provide theoretical proof that our proposed SGCL method has a high tolerance to noisy views. Further demonstration is given by conducting extensive experiments on three real-world datasets. The experimental results demonstrate that our approach substantially increases recommendation accuracy, with relative improvements reaching as high as 12.25% over nine other competing models. These results highlight the efficacy of our method.

Symmetric Graph Contrastive Learning against Noisy Views for Recommendation

TL;DR

This work tackles the problem that graph contrastive learning for recommendations can suffer when data augmentation creates noisy views that distort the original graph. It introduces Symmetric Graph Contrastive Learning (SGCL), a model-agnostic framework built on a symmetry-based loss that remains robust to noisy interference, and proves its compatibility with mutual information bounds via the Wasserstein Dependency Measure. The authors provide theoretical guarantees of noise tolerance and demonstrate that SGCL consistently improves recommendation accuracy on three real-world datasets, achieving up to 12.25% relative gains, while maintaining robustness under data sparsity and perturbations. These results offer a principled approach to robust self-supervised learning in graph-based recommender systems and suggest a practical, broadly applicable loss design for contrastive learning.

Abstract

Graph Contrastive Learning (GCL) leverages data augmentation techniques to produce contrasting views, enhancing the accuracy of recommendation systems through learning the consistency between contrastive views. However, existing augmentation methods, such as directly perturbing interaction graph (e.g., node/edge dropout), may interfere with the original connections and generate poor contrasting views, resulting in sub-optimal performance. In this paper, we define the views that share only a small amount of information with the original graph due to poor data augmentation as noisy views (i.e., the last 20% of the views with a cosine similarity value less than 0.1 to the original view). We demonstrate through detailed experiments that noisy views will significantly degrade recommendation performance. Further, we propose a model-agnostic Symmetric Graph Contrastive Learning (SGCL) method with theoretical guarantees to address this issue. Specifically, we introduce symmetry theory into graph contrastive learning, based on which we propose a symmetric form and contrast loss resistant to noisy interference. We provide theoretical proof that our proposed SGCL method has a high tolerance to noisy views. Further demonstration is given by conducting extensive experiments on three real-world datasets. The experimental results demonstrate that our approach substantially increases recommendation accuracy, with relative improvements reaching as high as 12.25% over nine other competing models. These results highlight the efficacy of our method.
Paper Structure (25 sections, 1 theorem, 17 equations, 8 figures, 8 tables, 1 algorithm)

This paper contains 25 sections, 1 theorem, 17 equations, 8 figures, 8 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Motivation
  3. Noisy Views in Data Augmentation
  4. Investigating the Impact of Noisy Views on Recommendation Accuracy
  5. Preliminaries
  6. Problem Definition
  7. Noise-tolerant Symmetric Loss
  8. Methodology
  9. Symmetric Form of Contrastive Learning
  10. Symmetric Contrastive Loss
  11. Multi-task Training
  12. Complexity Analysis
  13. Discussion
  14. Experiments
  15. Experiment settings
  16. ...and 10 more sections

Key Result

Corollary 1

Let $f^*_{\eta}=\mathrm{arginf}_{f \in \mathcal{F}}R^{\eta}_{\mathcal{L}}(f)$ be the minimizer of noisy risk, $f=\mathrm{arginf}_{f \in \mathcal{F}}R_{\mathcal{L}}(f)$ be the minimizer of optimal risk, and $\delta$ be $\mathbb{E}_{\mathcal{D}}[\mathcal{L}(f(x),y_x)]$, if the $\eta \le \eta_{max} < 0

Figures (8)

  • Figure 1: Overall framework of graph contrastive learning for collaborative filtering. It is worth mentioning that view 1 and view 2 are two generated contrastive views of $u_1$ (i.e., original graph). We observe that the structure of $u'_1$ is different from the original and $u"_1$, which hints that graph augmentation may introduce noisy views and thus degrade the final recommendation performance
  • Figure 2: Part (a) is the visualization of the embeddings of View 1 and the visualization of the embeddings of Original. Part (b) denotes the visualization of the embeddings of View 2 and the visualization of the embeddings of Original. Part (c) shows the number of noisy views in different level groups.
  • Figure 3: Model performance comparison with different noise ratios. We sequentially replace genuine user-item interactions with 5%, 10%, 15%, 20%, and 25% synthetic user-item edges.
  • Figure 4: Performance on the Yelp2020 and Amazon-Book datasets with varying degrees of sparsity level.
  • Figure 5: Performance v.s. Perturbation Rate. We increased the perturbation rate of edge dropout from 5% to 25%.
  • ...and 3 more figures

Theorems & Definitions (1)

  • Corollary