Predict what comes next
A causal Gaussian transformer estimates the next value and its uncertainty at every time step. It is trained without labels, calibrated, and then frozen.
Interpretable time-series · 2026
ConceptTime
Predictive Surprise as Self-Grounding Concept Bottleneck for Interpretable Time Series
A time-series classifier that discovers its own temporal concepts, explains decisions through them, and knows when an input falls outside what those concepts can describe.
17
benchmarks spanning wearable sensing, biomedicine, UEA datasets, and fault diagnosis
97.91%
Epilepsy accuracy with only 1% of the training labels
0.992
OOD AUROC on the strongest evaluated Gaussian-noise shift
01
Motivation
Interpretability starts before the classifier.
In healthcare, wearable sensing, and industrial monitoring, a useful model must be accurate, show which temporal patterns drove its decision, and defer when its evidence is unfamiliar. Existing time-series concept bottlenecks usually depend on fixed windows or expert-defined concepts.
The underlying problem is circular: meaningful concepts require meaningful segment boundaries, but meaningful boundaries depend on knowing what patterns to look for. Fixed windows can split a natural event, while raw-shape prototypes can treat two visually similar segments as equivalent even when their preceding context makes them behave very differently.
02
Method
One self-supervised signal, four connected stages.
Cut at surprise
Negative log-likelihood measures predictive surprise. A length-constrained dynamic program places a fixed number of boundaries where surprise is most informative.
Ground concepts
Each segment becomes a signature of predicted moments, standardized residual shape, and its surprise trajectory. A Gaussian mixture turns these signatures into inspectable prototypes.
Classify and abstain
A small transformer classifies the concept sequence. Distance to the nearest prototypes becomes a built-in fidelity score, with no extra uncertainty head.
The path from signal to concept activation is deterministic after self-supervised fitting. Only the lightweight classifier sees task labels, reducing the opportunity for hidden label information to leak into the concepts.
03
Performance
Competitive accuracy, especially when labels are scarce.
97.91%
Few-shot classification
On Epilepsy with 1% labels, ConceptTime exceeds every evaluated black-box and interpretable baseline; the best black-box comparison reaches 95.41%.
0.864→0.992
Unfamiliar-input detection
Fidelity AUROC rises smoothly as Gaussian noise severity increases from 0.5 to 2.0, while the comparison scores remain near chance.
86.99%
Cross-domain transfer
Average fault-diagnosis transfer accuracy with 5% target labels when the density model is adapted but the source concept vocabulary stays fixed.
| Benchmark | Full labels | 1% labels | 5% labels |
|---|---|---|---|
| UCI-HAR | 88.91% | 74.60% | 79.60% |
| Epilepsy | 98.39% | 97.91% | 98.01% |
| Sleep-EDF | 78.15% | 72.00% | 74.00% |
| UEA (10-dataset average) | 71.82% | 58.50% | 60.10% |
Results are means over three seeds. OOD figures use the HAR perturbation study; transfer figures average the 12 directed fault-condition pairs.
04
What the results show
Predictive geometry does more than adaptive patching alone.
Dynamic boundaries help, but the decisive signal is how a segment behaves relative to expectation. On identical adaptive patches, replacing predictive signatures with raw morphology causes large drops, including nearly 50 accuracy points on HAR in the ablation study.
The same prototype geometry also supports selective prediction and faithful patch ranking. This joins classification, interpretation, and abstention without annotations, language-model calls, or a separately trained calibration head.
- Fixed concept granularityThe vocabulary size is chosen rather than learned hierarchically.
- Human-readable, not automatically namedPrototypes are inspectable exemplars, but semantic labels still require human interpretation.
- Current scopeThe paper focuses on classification; online concept updates and anomaly localization remain future extensions.