Tag
The paper proposes XACT, a framework for learning sparse attribution masks over time-frequency transforms to explain time-series classifiers, demonstrating improved precision and interpretability over baselines.