A graph-based framework called G2I reported 100% coverage of all 51 at-risk individuals in a Military network, while its CF-Greedy method recorded substantially higher AUCC than CF and CF2 across intervention datasets. The coverage figure is a model-based result, not evidence that an intervention changed anyone's outcome.
The supplied document identifies itself as arXiv:2608.23835v1, dated 24 Aug 2026. Its stated question is whether simple, interpretable methods can generate good graph explanations, with counterfactual explanation reframed as an intervention-design problem.
Turning local explanations into a population policy
G2I starts with an iterative greedy local search over node features, edge modifications and perturbations to neighbor features. It then aggregates the local clauses into a global intervention policy, selecting them under a budget constraint.
In practical terms, the method moves from changes around one node and its neighborhood to a policy for population-level intervention design.
The evaluation used two kinds of data
The evaluation separates two jobs. Benchmark datasets with ground-truth motifs are used for explanation quality, while separate intervention datasets are used for counterfactual-based intervention design.
The intervention data include a Military network with 241 nodes and 258 edges, and Youth data on 131 individuals with 39 features and no explicit graph structure.
For prediction, the study used a unified graph-convolutional network, or GCN, with an 80/20 stratified train-test split. Reported average test accuracies were 95.85% for Military, 98.99% for Youth and 96.40% for synthetic datasets.
Each experiment was repeated 10 times with different random seeds. Means and standard deviations were reported, although the main tables showed means only.
What the comparisons found
On the explanation benchmarks, the CF-Greedy row reported precision entries of 78.87, 87.50 and 92.59. Its MIP entries were 1.20, 1.40 and 1.19, while the time-column entries were 0.037, 0.033 and 0.079.
For intervention comparisons, the authors reported substantially higher AUCC for CF-Greedy than for CF and CF2 across the intervention datasets.
The selection strategy also mattered. DNF was reported as best for CF2 and CF-Greedy when heterogeneous feature-edge clauses were available, while Frequency was often competitive or better for CF's edge-only explanations.
A separate diagnostic reported modularity ratios concentrated around 1, with almost all instances above 0.5, a pattern consistent with near-modularity in the tested setting.
A model-based blueprint
In the Military analysis, the framework reported 100% coverage of all 51 at-risk individuals. The figure is model-label coverage, not evidence that an intervention changed anyone's outcome.
The tiered intervention strategy was reported to cover 84% through career programs, 90% after clinical interventions and the remaining 10% through network-level strategies. These are model-based coverage percentages, not observed health outcomes.
Taken together, the figures describe candidate clauses selected under the framework's rules, not demonstrated human benefit.
A calculation that scales
In a separate scalability result, the greedy strategy scaled to graphs with over 20,000 nodes under the same memory constraints. That speaks to the size of the calculation, not to whether an intervention works outside the model analysis.
The work was supported by the Army Research Office grant W911NF-23-1-0354.
Paper data and sources
Original title: Generating Intervention Hypotheses using Explainable Explanations on Graphs: G2I, a Two-Stage Greedy Framework
Authors: Mulin Tian, Ajitesh Srivastava
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-24
DOI: 10.1145/3799682.3840645
Original paper · Full text