marginal-influence

This paper proves that existing marginal influence-based attribution methods fundamentally fail to capture the conditional dependency structure of time series models, and proposes DAG-faithfulness as a new criterion for faithful explanations.