DAG-SHAP feature attribution in directed acyclic graphs

DAG-SHAP, a new feature attribution method, was introduced in a paper submitted on 13 Jun 2026 by Qiheng Sun, Junxu Liu, Xiaokai Mao, Haocheng Xia, Jinfei Liu, Kui Ren and Haibo Hu (arXiv:2606.15273). The paper proposes treating edges in a directed acyclic graph as the unit of attribution and includes an approximation technique and experimental validation on real and synthetic datasets. The authors provide code at the URL linked in the paper.

What the paper introduces

The authors frame the problem around limitations of Shapley value-based feature attribution in settings with complex feature interactions and causal relationships. They state that existing methods typically adopt a node-centric view and therefore "often fail to simultaneously capture the externality and exogenous influence of features, leading to unreasonable interpretations." To address this, the paper presents DAG-SHAP, an attribution method "based on edge intervention." DAG-SHAP treats each feature edge as an individual attribution object, the authors write, so both externality and exogenous contributions are captured.

The submission includes an approximation method intended to compute DAG-SHAP efficiently. The paper reports extensive experiments on both real and synthetic datasets that, according to the authors, validate the effectiveness of DAG-SHAP. The arXiv entry lists the DOI 10.48550/arXiv.2606.15273 and provides links to PDF and source files.

How DAG-SHAP differs from node-centric Shapley approaches

Node-centric Shapley explanations allocate importance to individual features (nodes) without explicitly modelling the directed dependencies between features. The paper highlights two conceptual gaps in that approach: externality, meaning how a feature's effect propagates to others through edges, and exogenous influence, meaning contributions coming from outside the modeled node itself. DAG-SHAP reframes attribution around edges, making the directed influence that connects features the primary object of explanation.

The authors argue this edge-first view ensures that both forms of contribution are represented when computing attributions, and they pair the formulation with an approximation algorithm to make the computations tractable for practical datasets. The submission includes experimental comparisons on synthetic and real datasets; the paper describes those experiments as extensive, and the authors say the results validate DAG-SHAP's effectiveness.

Why it matters

Shapley-based methods are widely used because they provide a principled way to split contributions among features, but they can be brittle when causal structure and inter-feature influence are important. By moving the attribution target from nodes to edges, DAG-SHAP aligns the attribution object with the directed dependencies researchers and practitioners often care about in causal or structured data. If the approximation and experimental claims hold up under wider scrutiny, DAG-SHAP could offer clearer mechanistic interpretations in DAG settings where edge relationships drive outcomes.

What to watch

Look for independent replication using the code the authors link from the arXiv paper, and for follow-up studies that measure how the approximation method scales across larger DAGs and diverse real-world datasets. The paper's experimental validation on synthetic and real data is the immediate next milestone to inspect closely.