Simple and Asymmetric Graph Contrastive Learning without Augmentations

TL;DR

GraphACL proposes a simple, augmentation-free asymmetric contrastive objective for learning node representations that works on both homophilic and heterophilic graphs. By decoupling node identity and neighborhood context through two encoders and a predictor, and by predicting one-hop neighborhood signals (while encouraging two-hop monophily via an implicit alignment), GraphACL maximizes mutual information with neighborhood structure and provides downstream performance guarantees. The method achieves state-of-the-art results across 15 datasets, with especially large gains on heterophilic graphs, and ablations confirm the importance of the asymmetric architecture and uniformity term. This work highlights that simple, well-mounded objectives can outperform complex augmentation-based schemes in graph self-supervised learning, encouraging exploration of leaner strategies for robust graph representations.

Abstract

Graph Contrastive Learning (GCL) has shown superior performance in representation learning in graph-structured data. Despite their success, most existing GCL methods rely on prefabricated graph augmentation and homophily assumptions. Thus, they fail to generalize well to heterophilic graphs where connected nodes may have different class labels and dissimilar features. In this paper, we study the problem of conducting contrastive learning on homophilic and heterophilic graphs. We find that we can achieve promising performance simply by considering an asymmetric view of the neighboring nodes. The resulting simple algorithm, Asymmetric Contrastive Learning for Graphs (GraphACL), is easy to implement and does not rely on graph augmentations and homophily assumptions. We provide theoretical and empirical evidence that GraphACL can capture one-hop local neighborhood information and two-hop monophily similarity, which are both important for modeling heterophilic graphs. Experimental results show that the simple GraphACL significantly outperforms state-of-the-art graph contrastive learning and self-supervised learning methods on homophilic and heterophilic graphs. The code of GraphACL is available at https://github.com/tengxiao1/GraphACL.

Figures (15)

• Figure 1: The graphs of pure homophily and pure heterophily, where color denotes the semantic class. For both graphs, nodes with a similar one-hop neighborhood context have similar semantic classes and two-hop similarities still exist even without the one-hop homophily.
• Figure 2: An illustration of various design motivations. (a) The heterophilic graph where the color denotes node's semantic class. (b) Contrastive objectives with the homophily assumption encourage one-hop neighbors to have similar representations. GraphACL simply encourages the node to predict its neighbors, which can implicitly capture neighborhood context (c) and two-hop monophily (d).
• Figure 3: Illustration of existing contrastive schemes and GraphACL. (a) forces neighboring nodes to have similar representations based on the homophily assumption. (b) augments the graph and learns the augment-invariant representations of the same node. Our GraphACL in (c) simply reconstructs the neighborhood signal of each node based on an asymmetric predictor without relying on the homophily assumption and augmentation.
• Figure 4: The effect of representation dimension, and the pair-wise similarities of randomly sampled node pairs, one-hop and two-hop neighbors. More results are provided in Appendix \ref{['app:similarity']}.
• Figure 5: The testing performance varying decay rate $\lambda$. More results are given in Appendix \ref{['app:decay']}.

Theorems & Definitions (6)

• Theorem 1
• Theorem 2
• Theorem 3
• proof
• proof
• proof