Active Explanations Combat Shortcut Learning in GNNs

Graph Neural Networks (GNNs) are powerful, but they often suffer from "shortcut learning," where they exploit spurious correlations (edges, nodes, features) that are not causally related to the prediction. This compromises their reliability, especially when applied to out-of-distribution tasks. This research proposes XIGL, an architecture-agnostic, human-in-the-loop strategy designed to mitigate this problem. XIGL operates on two core insights. First, shortcut reliance can be identified by examining GNN explanations, which highlight the parts of the graph the model focuses on. Second, once these shortcuts are detected, expert users can provide targeted corrective feedback to deconfound the model, guiding it away from these non-causal features. While XIGL supports various query strategies, the authors developed an active learning approach to prioritize explanations most likely to reveal shortcut behavior. This active learning component reduces the annotation and cognitive costs for human experts. The effectiveness of XIGL, including both existing and newly proposed explanation-based strategies, is demonstrated across several GNN architectures.