Causal contrastive learning for generalizable graph classification

doi:10.1016/j.knosys.2026.116643

Causal contrastive learning for generalizable graph classification

2026-07-01

0

PRE

AI

DOI：10.1016/j.knosys.2026.116643

Original

Original request for help

Save

Abstract

En 中文

Graph classification is a fundamental task with broad real-world applications. Although graph neural networks (GNNs) can effectively model graph structures and node attributes, they often rely on spurious correlations between the input graphs and labels, which undermines robustness and generalization under distribution shifts. Recent studies attempt to address this issue by partitioning graphs into causal and shortcut subgraphs. However, such hard partitioning may disrupt the intrinsic graph structure and lead to the loss of important structural information. In addition, existing methods often focus on limited causal relations and thus fail to capture diverse semantic causal factors. To address these limitations, we propose Graph Causal Contrast Decoupling Learning (GCCDL), a framework for generalizable graph classification. Specifically, we design a multi-causal attention module to generate multi-mode graph views, enabling the model to capture distinct semantic causal features. We further incorporate a DropKey regularization mechanism into the attention process to improve structural robustness. Moreover, a contrastive learning objective is introduced to disentangle causal features from shortcut correlations by constructing positive and negative pairs across multi-mode graphs. Extensive experiments on six benchmark datasets demonstrate that GCCDL consistently achieves competitive and often superior performance compared with strong baselines.

Keywords:

Graph neural networks

Causal inference

Contrastive learning

Graph classification

Causal contrastive learning for generalizable graph classification

Abstract

Journal

Organization

Cited Papers