Adversarial Learning Data Augmentation for Graph Contrastive Learning in Recommendation

TL;DR

This paper introduces a novel adversarial data augmentation method guided by InfoMin and InfoMax principles for graph contrastive learning in recommendation systems, improving GNN performance on benchmark datasets.

Contribution

It proposes a new edge-operating data augmentation technique and a theoretical framework, LDA-GCL, that jointly optimize data augmentation and contrastive learning for better recommendations.

Findings

LDA-GCL outperforms existing methods on four benchmark datasets.

The adversarial augmentation improves the quality of learned representations.

The framework effectively balances information sharing and minimality in views.

Abstract

Recently, Graph Neural Networks (GNNs) achieve remarkable success in Recommendation. To reduce the influence of data sparsity, Graph Contrastive Learning (GCL) is adopted in GNN-based CF methods for enhancing performance. Most GCL methods consist of data augmentation and contrastive loss (e.g., InfoNCE). GCL methods construct the contrastive pairs by hand-crafted graph augmentations and maximize the agreement between different views of the same node compared to that of other nodes, which is known as the InfoMax principle. However, improper data augmentation will hinder the performance of GCL. InfoMin principle, that the good set of views shares minimal information and gives guidelines to design better data augmentation. In this paper, we first propose a new data augmentation (i.e., edge-operating including edge-adding and edge-dropping). Then, guided by InfoMin principle, we propose a novel theoretical guiding contrastive learning framework, named Learnable Data Augmentation for Graph Contrastive Learning (LDA-GCL). Our methods include data augmentation learning and graph contrastive learning, which follow the InfoMin and InfoMax principles, respectively. In implementation, our methods optimize the adversarial loss function to learn data augmentation and effective representations of users and items. Extensive experiments on four public benchmark datasets demonstrate the effectiveness of LDA-GCL.