[Paper Review] Explainable Artificial Intelligence (XAI) on TimeSeries Data: A Survey

Thomas Rojat, Raphaël Puget+4 more|arXiv (Cornell University)|Apr 2, 2021

Time Series Analysis and ForecastingComputer Science95 references118 citations

TL;DR

This survey critically reviews explainable AI methods applied to time series, categorizing techniques by model type and explanation scope, and discusses evaluation, trust, robustness, and interactive aspects.

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.

Motivation & Objective

• Provide an overview of XAI methods applied to time series and their methodology, scope, and targets.
• Discuss approaches that increase confidence, stability, and robustness of time series models.
• Evaluate qualitative and quantitative methods for explaining time-series models.
• Discuss limitations and potential future research directions in explainability for time series.

Proposed method

• Classify post-hoc explanations for CNNs (backpropagation-based and perturbation-based) and Ante-Hoc explanations for RNNs (attention mechanisms).
• Describe model-agnostic explanations (e.g., SHAP) applicable to recurrent models.
• Present data mining-based XAI methods specific to time series (SAX, fuzzy logic, rule-based systems, shapelets).
• Discuss explanations by example (prototype-based methods and shapelets) for time series.
• Differentiate local vs global explanations and discuss scalability and interpretability trade-offs.

Experimental results

Research questions

• RQ1 What XAI methods exist for time series, and how do they operate across different model families?
• RQ2 How do these methods provide explanations (local vs global) and what are their targets (developers vs end-users)?
• RQ3 What are the metrics and approaches to evaluate explanations in time series models?
• RQ4 What are the limitations and future directions for explainability in time series analytics?

Key findings

• CNN-based explanations include class activation mapping (CAM) and gradient-based methods to highlight time sub-sequences influential to predictions.
• Perturbation-based explanations like occlusion sensitivity identify important input segments by altering parts of the time series and observing output changes.
• RNNs can be explained with attention mechanisms, offering anticipatory or state-based interpretability embedded in the model design.
• Model-agnostic explanations (e.g., SHAP) can explain recurrent models, broadening applicability beyond CNNs/RNNs.
• Time-series specific explainability methods include SAX, fuzzy logic, and shapelets, enabling interpretable features and prototypes.
• The survey emphasizes trust, stability, robustness, and user-centered interactive explanations as essential yet underdeveloped aspects in time series XAI.