CAS Paper Reports 0.107 Mean Local MAE on Coalition-Aware Causal Benchmarks
CAS delivers a local-to-global causal reporting layer that outperforms predictive SHAP on treatment-effect modifier detection. Benchmarks confirm lower MAE under interaction regimes and materially different rankings on real policy datasets. The method supplies an explicit intervention target useful for security and privacy analysis of deployed models.
CAS converts interventional coalition games into Local CAS, Signed Local CAS and two Global CAS summaries using causal Shapley contributions. On known-truth data with n=2,200 per run and three actions, coalition-aware CAS reduced MAE from 0.173 (one-at-a-time) and 0.213 (global ATE). The gap widened to 0.091 under strong interactions versus near-zero under additivity.
Empirical tests on 401(k) eligibility data (n=9,915) and Pennsylvania reemployment data (n=5,099) showed Feature-CAS rankings diverging from TreeSHAP. In Pennsylvania, dep1 rose from predictive rank 13 to Feature-CAS rank 2, identifying the leading local modifier of unemployment-duration effects. This isolates heterogeneity drivers that standard predictive attributions miss.
Applied to browser-extension risk models, CAS could attribute causal impact of permission sets on data-leakage outcomes rather than mere correlation with usage signals. The explicit intervention target layer supplies a reporting structure that existing SHAP deployments lack for security audits.
Next steps include integration into DoubleML pipelines for extension telemetry and release of reference implementations on the two public datasets within six months.
DoubleML team: CAS integration into extension telemetry pipelines reaches production use in at least one browser vendor audit by Q3 2027.
Sources (2)
- [1]Primary Source(https://arxiv.org/abs/2608.12555)
- [2]Supporting Source(https://arxiv.org/abs/2206.06821)