CAS: A New Causal Attribution Score for Explainable AI
A recent study presents the Causal Attribution Score (CAS), a novel approach for causal explanations in artificial intelligence, now accessible on arXiv (2608.12555). This method interprets model outputs as the effects of interventions on actual outcomes, utilizing an interventional coalition game along with causal Shapley contributions. CAS produces Local CAS, Signed Local CAS, and two types of Global CAS summaries. In a benchmark involving eight repeated primary-interaction simulations (n = 2,200 each, three actions), CAS recorded a mean Local CAS MAE of 0.107, surpassing one-at-a-time normalization (0.173) and the global normalized absolute ATE vector (0.213). The research tackles challenges in explainable AI, improving reliability and clarity in understanding causal impacts. It falls under the category of ai-technology.
Key facts
- CAS is a compact score architecture for causal explanation.
- It starts from an identified interventional coalition game.
- It allocates joint intervention contrast with causal Shapley contributions.
- It converts raw outcome-scale effects into Local CAS, Signed Local CAS, and two Global CAS summaries.
- The innovation is a local-to-global causal reporting layer with an explicit intervention target.
- Benchmark: mean Local CAS MAE of 0.107 for coalition-aware CAS, 0.173 for one-at-a-time normalization, 0.213 for global normalized absolute ATE vector.
- Eight repeated primary-interaction simulations (n = 2,200 each, three actions) were used.
- The paper is available on arXiv (2608.12555).
Entities
Institutions
- arXiv