Feature Attribution in Directed Acyclic Graphs Using Edge Intervention
Researchers have introduced DAG-SHAP, a novel feature attribution method designed for directed acyclic graphs (DAGs) that addresses limitations of existing Shapley value-based approaches. Unlike previous node-centric methods that focus on individual features, DAG-SHAP treats feature edges as attribution objects to better capture externality and exogenous influence. The method includes an efficient approximation for computation, and experiments on synthetic and real datasets demonstrate its effectiveness. AI
IMPACT Introduces a more nuanced approach to feature attribution in causal AI models, potentially improving interpretability and reliability.