Concept Extraction for Time Series with ECLAD-ts

paper-conference
Concept extraction and localization for time-series CNNs, with gradient-based concept importance and validation on synthetic and natural datasets.
Authors

Antonia Holzapfel

Andrés Felipe Posada-Moreno

Sebastian Trimpe

Published

October 12, 2025

Doi

10.1007/978-3-032-08317-3_5

Resources

DOI | arXiv | PDF

Publication history

Preprint: 7 April 2025. Presented at xAI 2025 (9–11 July 2025). First published online by Springer: 12 October 2025. The publisher specifies 2026 as the citation year for CCIS volume 2576.

Summary

ECLAD-ts extends concept extraction to time-series CNNs by clustering timestep-wise descriptors assembled from activation maps at multiple layers. It preserves temporal and channel information, localizes the resulting concepts, and measures their influence on predictions using model gradients. Experiments on synthetic and natural datasets compare the method with existing approaches and examine the strengths and limitations of concept-based explanations for time series.

Citation

@inproceedings{holzapfel2026ecladts,
 author = {Holzapfel, Antonia and Posada-Moreno, Andrés Felipe and Trimpe, Sebastian},
 booktitle = {Explainable Artificial Intelligence},
 date = {2026},
 doi = {10.1007/978-3-032-08317-3_5},
 pages = {90--112},
 publisher = {Springer},
 series = {Communications in Computer and Information Science},
 title = {Concept Extraction for Time Series with ECLAD-ts},
 volume = {2576}
}