GPS: Graph Contrastive Learning via Multi-scale Augmented Views from Adversarial Pooling

TL;DR

The paper tackles the challenge of self-supervised graph representation learning by addressing the reliance on handcrafted augmentations. It introduces GPS, which uses learnable graph pooling to automatically generate two multi-scale augmented views—one strongly augmented and one weakly augmented—and integrates them in a joint contrastive framework with adversarial training between the pooling modules and the encoder. GPS combines a hard similarity objective for weak augmentations and a soft distributional consistency objective for strong augmentations, enabling robust and semantically meaningful representations. Extensive experiments across graph classification, transfer learning, clustering, and semi-supervised tasks on twelve datasets demonstrate GPS's superior performance and robustness, highlighting the practical value of learnable, adversarially trained multi-scale augmentations in graph contrastive learning.

Abstract

Self-supervised graph representation learning has recently shown considerable promise in a range of fields, including bioinformatics and social networks. A large number of graph contrastive learning approaches have shown promising performance for representation learning on graphs, which train models by maximizing agreement between original graphs and their augmented views (i.e., positive views). Unfortunately, these methods usually involve pre-defined augmentation strategies based on the knowledge of human experts. Moreover, these strategies may fail to generate challenging positive views to provide sufficient supervision signals. In this paper, we present a novel approach named Graph Pooling ContraSt (GPS) to address these issues. Motivated by the fact that graph pooling can adaptively coarsen the graph with the removal of redundancy, we rethink graph pooling and leverage it to automatically generate multi-scale positive views with varying emphasis on providing challenging positives and preserving semantics, i.e., strongly-augmented view and weakly-augmented view. Then, we incorporate both views into a joint contrastive learning framework with similarity learning and consistency learning, where our pooling module is adversarially trained with respect to the encoder for adversarial robustness. Experiments on twelve datasets on both graph classification and transfer learning tasks verify the superiority of the proposed method over its counterparts.

Figures (5)

• Figure 1: Illustration of the proposed framework GPS. We first generate two positive views via our two pooling modules. Then, the two augmented views are fed into the online network while the original graph is fed into the target network. Our contrastive learning framework captures similarity learning and consistency learning, where the graph pooling modules are adversarially trained with respect to the encoder.
• Figure 2: Illustration of the graph pooling methods.
• Figure 3: Performance of ablation study of several model variants (in $\%$) on all six datasets.
• Figure 4: Analysis of graph pooling ratio on IMDB-B.
• Figure 5: Analysis of batch size on PROTEINS and IMDB-B.