Evaluation of post-hoc interpretability methods in time-series classification

Abstract Post-hoc interpretability methods are critical tools to explain neural-network results. Several post-hoc have emerged in recent years but they produce different results when applied a given task, raising the question of which method is most suitable provide accurate interpretability. To understand performance each method, quantitative evaluation essential; however, currently available frameworks several drawbacks that hinder adoption methods, especially high-risk sectors. In this work we propose framework with metrics assess existing particularly time-series classification. We show identified literature addressed, namely, dependence on human judgement, retraining and shift data distribution occluding samples. also design synthetic dataset known discriminative features tunable complexity. The proposed methodology can be used reliability obtained practical applications. turn, embedded within operational workflows fields require for, example, regulatory policies.