DAG-SHAP Improves Feature Attribution in Causal Models

Traditional Shapley value-based feature attribution methods often struggle with complex feature interactions and causal relationships, even when a causal structure is explicitly provided. These methods typically adopt a node-centric perspective, assigning importance solely to individual features. This can lead to incomplete or misleading interpretations, as they often fail to simultaneously account for both the external impact and the independent, exogenous influence of features within a causal graph. To address these shortcomings, researchers have developed DAG-SHAP, a novel feature attribution method grounded in edge intervention. Instead of focusing on individual features (nodes), DAG-SHAP treats each feature edge within a Directed Acyclic Graph (DAG) as a distinct attribution object. This allows the method to appropriately capture both the externality and exogenous contributions of features. The paper also introduces an efficient approximation method for computing DAG-SHAP. Extensive experiments on both synthetic and real-world datasets confirm the effectiveness of DAG-SHAP in providing more accurate and insightful feature attributions compared to existing techniques.