ConceptTime uses predictive surprise to interpret time-series classifiers
Original title: Predictive Surprise as Self-Grounding Concept Bottleneck for Interpretable Time Series
A frozen forecaster's prediction error both segments series and builds concepts, with no labels, language model, or expert vocabulary.
Researchers introduce ConceptTime, a self-supervised concept-bottleneck method for risk-sensitive time-series classification. A frozen probabilistic forecaster scores mismatch with the next observation, called predictive surprise, to place segment boundaries and summarize dynamical regimes as concepts. A lightweight head classifies from the concept sequence, and distance to the nearest concept supplies reliability. Across 17 public benchmarks the paper reports accuracy competitive with black-box models, better than interpretable baselines, and near-98% on Epilepsy with 1% of labels, without annotations.
Unconfirmed: Benchmark claims are the authors' reported results, not independently verified here.
- Predictive surprise is the mismatch between a frozen forecaster and the realized signal.
- Near-98% accuracy is reported on Epilepsy using 1% of labels.
- Evaluated against softmax, prototype-distance, and SHAP/TimeX/TimeX++ for OOD and deletion faithfulness.
- Code is at github.com/Sachithx/ConceptTime.git.
From the original article
1 min read at huggingface.co
Time-series classifiers deployed in risk-sensitive domains must be (1) accurate, (2) interpretable, and (3) able to abstain when evidence is weak. Existing methods do not deliver all 3 at once, and the obstacle is structural, rather than logistical. Meaningful temporal patterns cannot be defined without knowing segment boundaries, yet meaningful boundaries cannot be placed without knowing which patterns to expect.
By Sachith Abeywickrama, Emadeldeen Eldele, Min Wu · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.