A research team at Zhejiang University has proposed a new way to explain machine learning predictions that shifts focus from individual features to the relationships between them.
The paper introduces DAG-SHAP, a variant of Shapley value attribution designed for models where features are connected through a directed acyclic graph (DAG) — a common structure in causal modeling. Standard SHAP methods assign importance scores to each feature in isolation. DAG-SHAP instead assigns scores to each edge in the graph, the links between features, arguing that this better captures how features influence each other and the model's output. The team also published an approximation algorithm, since computing exact Shapley values scales exponentially with feature count. Experiments on synthetic and real-world datasets are offered as validation, and the code is on GitHub.
Shapley values have become the dominant explainability framework in applied ML, but they carry a well-known assumption: features contribute to a prediction more or less independently. That assumption strains when working with causal graphs, where feature A's effect on an output partly runs through feature B. Node-level attribution in that setting can overstate or understate a feature's importance depending on where it sits in the causal chain. An edge-first view could give practitioners a clearer read on which causal pathways actually drive a prediction, not just which variables show up.
The approximation method will be the real test. Exact Shapley computation is exponential; most production uses of SHAP already depend on approximations like TreeSHAP or KernelSHAP. Whether DAG-SHAP's version is fast enough for large causal graphs — and faithful enough not to introduce its own distortions — is a question the published experiments alone won't settle.