Adversarial Curriculum Graph Contrastive Learning with Pair-wise Augmentation

TL;DR

This work tackles the challenge of generating semantically meaningful positive and negative samples for graph contrastive learning by introducing Adversarial Curriculum Graph Contrastive Learning (ACGCL). The method pairs a novel pair-wise graph augmentation that yields mirror graphs with controllable similarity (via a distance threshold ${\gamma}$) with a subgraph contrastive learning framework that concurrently learns inter-graph and intra-graph patterns; it is further enhanced by an adversarial curriculum learning (ACL) strategy that reweights samples according to their difficulty using a max–min optimization over weights ${u_i}$ and ${v_i}$. The key contributions are the mirror-graph augmentation mechanism, the subgraph-level contrastive objective with a Wasserstein-based balance term, and the ACL framework that mitigates sample-sparsity issues while progressively increasing training difficulty. Empirical results on six benchmark datasets show state-of-the-art performance for node classification, demonstrating improved robustness and scalability over prior graph contrastive approaches. This work advances graph representation learning by enabling fine-grained control over positive/negative similarity and by efficiently leveraging subgraph structures under adversarial curriculum cues.

Abstract

Graph contrastive learning (GCL) has emerged as a pivotal technique in the domain of graph representation learning. A crucial aspect of effective GCL is the caliber of generated positive and negative samples, which is intrinsically dictated by their resemblance to the original data. Nevertheless, precise control over similarity during sample generation presents a formidable challenge, often impeding the effective discovery of representative graph patterns. To address this challenge, we propose an innovative framework: Adversarial Curriculum Graph Contrastive Learning (ACGCL), which capitalizes on the merits of pair-wise augmentation to engender graph-level positive and negative samples with controllable similarity, alongside subgraph contrastive learning to discern effective graph patterns therein. Within the ACGCL framework, we have devised a novel adversarial curriculum training methodology that facilitates progressive learning by sequentially increasing the difficulty of distinguishing the generated samples. Notably, this approach transcends the prevalent sparsity issue inherent in conventional curriculum learning strategies by adaptively concentrating on more challenging training data. Finally, a comprehensive assessment of ACGCL is conducted through extensive experiments on six well-known benchmark datasets, wherein ACGCL conspicuously surpasses a set of state-of-the-art baselines.

Figures (5)

• Figure 1: Overview of the proposed framework ACGCL
• Figure 1: Parameter analysis
• Figure 2: An example of one step in pair-wise graph augmentation. Let a solid (dashed) edge denote that there is a (no) edge between the two nodes. Given a pair of nodes $v_a$ and $v_b$ in the original graph, their positive mirror pair $v_c$ and $v_d$ and negative mirror pair $v_e$ and $v_f$, the green edge between node $v_{a}$ and $v_{b}$ in the original graph is maintained to create the positive mirror graph as the positive mirror pair $v_c$ and $v_d$ are connected. While it will be removed in the negative mirror graph as there is no edge between the pair $v_e$ and $v_f$.
• Figure 2: Visualization of subgraph embedding distribution
• Figure 3: Graph sample difficulty validation results. With larger $\theta$, the sample difficulty of mirror subgraphs from pair-wise graph augmentation increases.

Theorems & Definitions (6)

• Theorem 1
• Proposition 1
• Proposition 2
• Theorem 1
• Proposition 1
• Proposition 2