A paper on arXiv introduces DAG‑SHAP, a feature‑attribution technique that works on directed acyclic graphs by intervening on edges rather than nodes.
The authors argue that existing Shapley‑based methods assume a node‑centric view, which can ignore how features influence each other through causal links. DAG‑SHAP assigns importance to each edge, thereby recording both the direct contribution of a feature and the spill‑over effects it creates. They also present an approximation algorithm to keep computation tractable. Experiments on synthetic benchmarks and real‑world datasets show that DAG‑SHAP produces more sensible attributions than conventional node‑based SHAP variants, and the code is publicly released on GitHub.
If the claim holds, DAG‑SHAP could narrow a known gap in explainable AI: representing causal structure without collapsing it into isolated feature scores. Practitioners who already use SHAP for model debugging may find edge‑level granularity useful for complex pipelines such as recommendation systems or gene‑expression networks, where interactions matter as much as individual signals.
The work is a reminder that many attribution tools still treat causal graphs as afterthoughts; DAG‑SHAP pushes the conversation toward truly causal explainability, but its utility will depend on how often users can supply reliable DAGs.