Explainable Artificial Intelligence (XAI) on TimeSeries Data: A Survey

TL;DR

This survey addresses the interpretability gap of deep learning on time-series by cataloging explainable AI methods across CNNs, RNNs, and data-mining approaches. It distinguishes post-hoc versus ante-hoc explanations, local versus global scales, and audience-specific needs, while detailing how explanations relate to stability, robustness, and confidence. The paper highlights evaluation challenges in time-series XAI and advocates for interactive, user-centered designs and metrics that quantify trust and robustness. Overall, it emphasizes that advancing time-series XAI requires domain-aware methods, end-to-end explainability systems, and rigorous, multi-faceted evaluation to support real-world deployment in critical areas like healthcare and autonomous systems.

Abstract

Most of state of the art methods applied on time series consist of deep learning methods that are too complex to be interpreted. This lack of interpretability is a major drawback, as several applications in the real world are critical tasks, such as the medical field or the autonomous driving field. The explainability of models applied on time series has not gather much attention compared to the computer vision or the natural language processing fields. In this paper, we present an overview of existing explainable AI (XAI) methods applied on time series and illustrate the type of explanations they produce. We also provide a reflection on the impact of these explanation methods to provide confidence and trust in the AI systems.

Figures (9)

• Figure 1: Questions answered throughout the survey and their connection to the different sections in which it is structured.
• Figure 2: Knowledge graph relating all the purposes of explainability methods for time series.
• Figure 3: Rationale connecting the contents of Section \ref{['Terminology']}.
• Figure 4: Usage of 'Gradient*Input' to identify the contribution of the input raw data when performing time series classification. It extracts the highly activated nodes in a channel and visualize the input sub-sequences that contribute to the highly activated nodes. Then, each extracted sub-sequence is assigned to a cluster of similar patterns. Figure reproduced with authorization from Cho et al. cho2020interpretation.
• Figure 5: (a) Sample time series with top-2 relevant filters from Two Patterns dataset, (b) their activation maps, and (c) occlusion sensitivity plot. The goal is to use the occlusion sensitivity method zeiler2013visualizing to compute the raw input contribution. Figure reproduced with authorization from Kashiparekh et al. Kashiparekh2019.