G2I Generates Actionable Intervention Hypotheses on Graphs

This research introduces G2I, a novel two-stage greedy framework designed to generate actionable intervention hypotheses from Graph Neural Networks (GNNs). While GNNs excel at modeling relational data, existing explanation methods often fall short in providing practical, network-level intervention strategies, particularly for complex real-world decision-making in fields like public health. The proposed framework redefines the problem of counterfactual explanation as an intervention design challenge. In its first stage, it employs a greedy search at the local level to identify the minimal and most actionable changes to node features and neighbor conditions that would alter a GNN's prediction. These counterfactuals are then translated into easily interpretable rules, making them accessible to domain specialists who may not have AI expertise. The second stage addresses network-level intervention selection, formulating it as a Disjunctive Normal Form (DNF) coverage problem with budget constraints. This allows for a greedy algorithm with theoretical guarantees to select cost-effective intervention strategies. Experiments on synthetic graphs and real-world suicide risk networks demonstrated that G2I produces scalable and efficient intervention strategies, significantly outperforming mask-based counterfactual methods in terms of efficiency.