Unraveling and Mitigating Endogenous Task-oriented Spurious Correlations in Ego-graphs via Automated Counterfactual Contrastive Learning

Graph Neural Networks (GNNs) have been proven to easily overfit spurious subgraphs in the available data, which reduces their trustworthiness in high-stakes real-world applications. Current works often assumed that such spurious subgraphs are caused by a latent environment variable (e.g., selection bias) and addressed this issue by learning invariance across synthesized multiple environments. However, this work uncovers a prevalent yet overlooked mechanism in node-level tasks leading to spurious subgraphs without assumption on the environment variable. Moreover, the identified mechanism implies that the spurious subgraphs can differ indifferent tasks even within the same ego-graph. To mitigate this E ndogenous and T ask-oriented S purious C orrelations (ETSC), this work designs a novel and automated C ounterfactual C ontrastive L earning framework for G raphs in n ode-level tasks (CCL-Gn). Based on the analysis of the relationship between spurious subgraphs and causally correlated subgraphs to the task within this mechanism, we propose an original counterfactual optimization objective to separate them automatically and sufficiently. To further maintain a model-agnostic property, CCL-Gn enables GNN optimization with an auxiliary contrastive learning objective between the raw ego-graph and the counterfactual views. Extensive experiments on 13 datasets with 29 data splits demonstrate that CCL-Gn can consistently enhance the performance of a series of typical GNNs in both the in-distribution and out-of-distribution scenarios.