CENDRe: Concept Extraction with Natural Domain Representations
A paper titled 'CENDRe: Concept Extraction with Natural Domain Representations' was published on arXiv on 2026-07-31. It proposes a concept extraction method for CNNs used in time-series classification, addressing limitations of existing methods by operating in both time and frequency domains, automatically selecting the number of concepts via silhouette-guided aggregation, and producing localizations aligned with model-used regions through gradient-based masks.
Researchers introduced CENDRe, a concept extraction method for convolutional neural networks applied to time-series classification. The method discovers concepts by clustering per-timestep latent representations in two stages, automatically determining the number of concepts using silhouette-guided aggregation. It localizes concepts by computing gradients of a presence score that contrasts latent representations with prototypes, generating masks focused on relevant regions. By propagating gradients through a differentiable invertible mapping like a Fourier transform, CENDRe provides localizations in both time and frequency domains, overcoming limitations of prior methods that only operated in the time domain, required predefined concept counts, and produced misaligned localizations.
CENDRe leverages a two-stage clustering of per-timestep latent representations with silhouette-guided aggregation to automatically determine the number of concepts, avoiding manual specification. Concept localization is achieved via gradient-based masks derived from a presence score that contrasts latent representations with prototypes, and by using a differentiable invertible mapping (e.g., Fourier transform), the method extends localization to the frequency domain, enabling joint time-frequency concept extraction.
This method enhances interpretability of CNN-based time-series classifiers, which are critical in domains like healthcare, finance, and industrial monitoring. By providing more accurate and automatic concept extraction in both time and frequency domains, CENDRe could improve trust and debugging in deployed models, potentially accelerating adoption in regulated industries where explainability is mandatory.
CENDRe addresses a key barrier to deploying CNNs in high-stakes time-series applications by providing interpretable, automatically discovered concepts. This could reduce the cost and effort of model validation, support compliance with explainability regulations, and enable new products in sectors like predictive maintenance, medical diagnostics, and financial forecasting.
Next signals include empirical validation on real-world time-series datasets, comparison with state-of-the-art concept extraction methods, and potential integration into explainability toolkits. Further research may explore extending the approach to other neural architectures or modalities, and assessing its impact on model debugging and regulatory compliance.