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

TL;DR

This work introduces a model-agnostic framework to quantitatively evaluate post-hoc interpretability methods for time-series classification, addressing key challenges such as dependence on human judgement, retraining, and distribution shift during occlusion. It defines two metrics, $AUC\tilde{S}_{\textrm{top}}$ and $F1\tilde{S}$, to assess relevance identification and supports their interpretation with TIC-based attribution curves, all while avoiding retraining. A new tunable synthetic, multivariate dataset complements experiments on Ford A and ECG across CNN, Bi-LSTM, and Transformer architectures, yielding robust rankings of interpretability methods. Across datasets, Shapley Value Sampling consistently yields the best relevance identification, with Integrated Gradients and DeepLiftShap excelling in specific settings, providing practical guidance for deploying reliable post-hoc explanations in high-stakes domains. The work also provides a theoretical link between attribution and information content, and releases code within the InterpretTime framework to foster adoption and benchmarking in real-world workflows.

Abstract

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

Figures (19)

• Figure 1: Relevance produced by four post-hoc interpretability methods, obtained on a time-series classification task, where a Transformer neural network needs to identify the pathology of a patient from ECG data. Depicted in black are two signals, V1, and V2, while the contour map represents the relevance produced by the interpretability method. Red indicates positive relevance, while blue indicate negative relevance. The former marks portions of the time series that were deemed important by the interpretability method for the neural-network prediction. The latter marks portions of the time series that were going against the prediction.
• Figure 1: $\tilde{S}$ as a function of the ratio of points removed with respect to the total number of time steps in the sample, $\tilde{N}$. Each subfigure represents one of the six interpretability methods considered for a Bi-LSTM trained on the synthetic dataset.
• Figure 1: $\tilde{S}$ as a function of the ratio of points removed with respect to the total number of time steps in the sample, $\tilde{N}$. Each subfigure represents one of the six interpretability methods considered for a Bi-LSTM trained on the Ford A dataset.
• Figure 2: Sample from the synthetic dataset.
• Figure 2: $\tilde{S}$ as a function of the ratio of points removed with respect to the total number of time steps in the sample, $\tilde{N}$. Each subfigure represents one of the six interpretability methods considered for a CNN trained on the synthetic dataset.