Feature Attribution in Directed Acyclic Graphs Using Edge Intervention

A team of researchers has proposed DAG-SHAP, a feature attribution method for directed acyclic graphs that shifts the unit of analysis from individual nodes to feature edges, aiming to better capture both externality and exogenous contributions in causal models [1]. The method, detailed in a paper submitted to arXiv on 13 June 2026, addresses a known shortcoming of Shapley value-based attribution techniques. Those methods typically adopt a node-centric view, attributing importance solely to individual features, and often fail to simultaneously capture the externality and exogenous influence of features, leading to what the authors call unreasonable interpretations [2]. DAG-SHAP instead treats each feature edge as an individual attribution object [2]. The authors also introduce an approximation method for efficiently computing DAG-SHAP and validate its effectiveness through experiments on both real and synthetic datasets [2]. The code has been made publicly available on GitHub [2]. The paper appears on arXiv, an open-access repository of electronic preprints that, as of November 2024, receives about 24,000 submissions per month and has surpassed two million total articles [6]. The work is listed under the Computer Science > Artificial Intelligence category and includes links to several community-developed tools integrated through arXivLabs, a framework launched by arXiv to allow collaborators to build experimental features directly on the platform [4][1]. arXivLabs projects, which include the Bibliographic Explorer and CORE Recommender, are developed under guidelines that require partners to share arXiv’s values of openness, community, excellence, and user data privacy [4][5]. The arXivLabs framework is currently on a temporary hiatus for new proposals while the development team focuses on modernizing and migrating arXiv’s systems to the cloud, though existing projects and already-submitted proposals are unaffected [3]. The DAG-SHAP paper’s abstract page also surfaces links to code-finding services such as CatalyzeX Code Finder for Papers and Papers with Code, which aim to provide an easy way to find relevant code for machine learning papers [1][5].