Implet: A Post-hoc Subsequence Explainer for Time Series Models

TL;DR

Implet addresses explainability in time series by post-hoc extracting salient subsequences from feature attributions. The method formalizes a subsequence score $s(l,r; x,w) = \left(\sum_{i=l}^{r} |w_i|\right) + \lambda (r-l+1)$ and enforces length bounds to produce non-overlapping implets; Coh-Implet clusters these into centroid explanations using a 2D DTW distance $d_{dtw}$ and DBA to summarize cohorts. Empirical evaluations on 13 UCR datasets with FCN and InceptionTime show that Implet explanations are faithful to the model and more concise than baselines, with Coh-Implet centroids preserving predictive factors. The code is open-sourced and the approach generalizes to higher-dimensional data. The work advances time-series XAI by delivering a scalable post-hoc, subsequence-level explanation framework and releasing open-source code.

Abstract

Explainability in time series models is crucial for fostering trust, facilitating debugging, and ensuring interpretability in real-world applications. In this work, we introduce Implet, a novel post-hoc explainer that generates accurate and concise subsequence-level explanations for time series models. Our approach identifies critical temporal segments that significantly contribute to the model's predictions, providing enhanced interpretability beyond traditional feature-attribution methods. Based on it, we propose a cohort-based (group-level) explanation framework designed to further improve the conciseness and interpretability of our explanations. We evaluate Implet on several standard time-series classification benchmarks, demonstrating its effectiveness in improving interpretability. The code is available at https://github.com/LbzSteven/implet

Figures (8)

• Figure 1: The workflow of the proposed post-hoc subsequence explainer Implet. It first computes attributions, then identifies continuous subsequences with consistently high attributions. Lastly, subsequences with similar shapes and attributions are clustered together to enhance conciseness.
• Figure 2: Implet cohort explanations for the GunPoint dataset. Each subfigure corresponds to one cohort, with highlighted regions representing the identified implets. The bold subsequence to the right of the dividing line denotes the Coh-Implet (cluster centroid) The color intensity represents attribution strength.
• Figure 3: Implet cohort explanations for the Chinatown dataset. Each figure represents one cohort, with highlighted regions denoting the identified implets. The bold subsequence to the right of the dividing line denotes the Coh-Implet (cluster centroid). The color intensity represents attribution strength.
• Figure 4: Example of the proposed subsequence removal scheme.
• Figure 5: Faithfulness evaluation of implets with different attribution methods, compared against the baseline ShapeletTransform. Arrows indicate the accuracy drop from removing identified subsequences (arrow tip) versus removing random subsequences of equal length (arrow tail). e.g. an arrow pointing from 0.2 to 0.4 represents removing random subsequence causes an accuracy drop of 0.2, while removing the explainer output causes an accuracy drop of 0.4. Longer, upward arrows indicate more faithful explanations. Short horizontal lines indicate negligible differences between random and identified subsequence removal. Dashed arrow represent ShapeletTransform for clarity. Horizontal dashed lines correspond to the accuracy drops that are equivalent to random guess. The last two dataset differ from the rest as Earthquaks is event-based and FordA is frequency-based.

Theorems & Definitions (1)

• Definition 1