ConceptTime turns forecast error into segment boundaries, keeping 97.91% epilepsy accuracy with 1% of labels

Time-series classifiers can now expose inspectable concepts, competitive accuracy and a reliability score for abstention at once, without predefined windows, expert concept labels or language-model calls. Earlier concept-bottleneck models either paired fixed windows with predefined concept vocabularies, segmented adaptively but fed patches into ordinary encoder pipelines, or produced discrete latent codes with no semantic structure, giving up at least one of accuracy, interpretability or calibrated abstention. Across 17 public benchmarks, ConceptTime-Dynamic is the strongest interpretable method at every label fraction; on Epilepsy it reaches 98.39% with full supervision and 97.91% with 1% of labels, above the black-box TSLANet at 95.41%.

Segmentation and concept description come from one frozen probabilistic forecaster: at each step it predicts a Gaussian over the next observation, and the negative log-likelihood of the realized signal under that prediction is the predictive surprise. High surprise means the model was wrong despite being confident, which is exactly where a boundary belongs; the same forecaster's predicted means, variances and residual statistics describe how each segment behaved relative to expectation rather than what it looks like. On UCI-HAR, with the same adaptive patches and vocabulary size, predictive-moment features reach 87.7% accuracy versus 42.7% for raw-signal morphology, a 45.0-point difference.

The path from raw signal to concept activation is fully deterministic: surprise is segmented by length-constrained dynamic programming, each patch is reduced to predictive moments, residual morphology and surprise-trajectory signatures, and a Gaussian mixture over signatures yields concept prototypes. Only a lightweight classifier on the concept tensor is trained with labels, while the frozen forecaster, standardizer and prototypes are shared across tasks, so concepts cannot smuggle label information into nominally interpretable activations. Across the 12 transfer directions among fault-diagnosis conditions FD-A to FD-D, keeping the source vocabulary and training only the classifier with 5% of target labels averages 91.09% accuracy; retraining the forecaster on unlabeled target inputs raises this to 94.37%, ranking first on 7 directions.

The demonstrated scope is 17 public benchmarks spanning wearable sensing, EEG seizure and sleep staging, 10 UEA datasets and four bearing fault conditions; the abstention score is evaluated on synthetic noise and shuffle perturbations on HAR and on fault-diagnosis condition shifts. The authors propose future work on hierarchical vocabularies, online vocabulary updates, and using low-fidelity segments to guide active relabeling.

Vocabulary size must be fixed in advance, and the meaning of each learned prototype depends on human interpretation after training; in fault-diagnosis condition transfer, fidelity detection averages AUROC 0.666 but still has a mean false-positive rate of 0.719 at 95% OOD recall, with large variation across condition pairs; on Epilepsy selective prediction, max softmax has lower E-AURC than fidelity; the authors also caution that high fidelity is not evidence of correctness, since a prototype may reflect a sensor artifact.