Robust Relatable Explanations of Machine Learning with Disentangled Cue-specific Saliency

DOI

Concept-based explanations help users understand the relation between model predictions and meaningful cues. However, under noisy real-world conditions, data perturbations lead to distorted and deviated explanations. We hypothesize that these corruptions affect specific cues rather than all, so disentangling them may help reduce model dependency on degraded cues. For the application of explaining emotional speech recognition, we propose RobustRexNet to explain with disentangled and discretized saliency maps for separate speech cues (e.g., loudness, pitch) to improve robustness against noise. Modeling evaluations show that RobustRexNet improved both model performance and explanation faithfulness in noisy and privacy-preserving distortions. User studies further indicate that the robust explanations aligned better with human intuition and improved user emotion labeling under noise. This work contributes toward robust explainable AI to improve user trust under real-world conditions.