🧩 ConceptTime: Datasets & Checkpoints

Predictive Surprise as Self-Grounding Concept Bottleneck for Interpretable Time Series (NeurIPS 2026)

Sachith Abeywickrama, Emadeldeen Eldele, Min Wu, Xiaoli Li, Chau Yuen

This repository holds the data and trained checkpoints that go with the ConceptTime code on GitHub. There are two parts:

• Ready-to-evaluate assets for UCI-HAR and Sleep-EDF: the test split plus both trained checkpoints. download_data.py fetches them to the exact paths the evaluation script expects. Epilepsy, the third headline dataset, ships directly in the GitHub repo.
• An archive/ with all 17 benchmarks (full train/val/test splits) and all trained checkpoints, including ablations and sweeps. Use it to retrain, run ablations, or explore beyond the headline results.

What is ConceptTime?

Interpretable time-series classifiers face a chicken-and-egg problem. You can't define meaningful temporal concepts without knowing where segments begin and end, and you can't place meaningful boundaries without knowing which concepts to look for. ConceptTime breaks this loop with a single self-supervised signal, predictive surprise.

1. A frozen probabilistic density model (a causal transformer) predicts a distribution over the next value at every timestep.
2. The mismatch between that prediction and the actual value is the surprise. It places segment boundaries, and it describes each segment by how it behaved relative to expectation rather than by what it looks like.
3. The segment summaries cluster into a vocabulary of dynamical regimes, called concepts. A lightweight transformer head then classifies from the sequence of concepts.
4. The distance to the nearest concept gives a reliability score by design, which supports calibrated abstention and OOD detection.

This needs no human concept annotations, no language model calls, and no domain experts. Across 17 public benchmarks, ConceptTime is competitive with state-of-the-art black-box classifiers, beats every interpretable baseline, and keeps about 98% accuracy on Epilepsy with only 1% of the labels.

Repository contents

dataset/                       test splits for the headline datasets
├── HAR/test.pt
└── SLeep-EDF/test.pt
checkpoints/                   trained checkpoints for the headline datasets
├── HAR/{density_model_best,pipeline_best}.pt
└── SLeep-EDF/{density_model_best,pipeline_best}.pt
archive/                       everything (~4.2 GB)
├── datasets/<Dataset>/{train,val,test}.pt          (FD-X: {train,val,test}_x.pt)
└── checkpoints/<Dataset>/
    ├── gaussian_entropy_best.pt                    density model
    ├── pipeline/entropy/pipeline_M<K>_*.pt         ConceptTime-Dynamic (surprise-based segmentation)
    ├── pipeline/static/pipeline_M<K>_*.pt          ConceptTime-Static (fixed-window ablation)
    └── sweep/...                                    hyperparameter sweeps (some datasets)

Each dataset needs two checkpoints for evaluation:

• density_model_best.pt is the frozen density model. It drives segmentation and the surprise signatures. In archive/ this file is called gaussian_entropy_best.pt, its older name.
• pipeline_best.pt holds the concept space, the signature standardizer, and the classifier.

The archive pipeline filenames record their configuration. For example, in pipeline_M32_surprise_surprise_full_gmm_transformer_..._ls0.6_lt0.5.pt, M32 is a vocabulary of 32 concepts, the first surprise is the boundary mode, surprise_full is the signature family, gmm is the clustering method, transformer is the classifier head, and the trailing values are loss weights.

Data format. Every .pt file is a dictionary:

{"samples": FloatTensor[N, C, T], "labels": LongTensor[N]}

Datasets

Dataset	Domain	Train / Test	Channels	Length T	Classes
UCI-HAR	Wearable inertial sensing	7,352 / 2,947	9	128	6 activities
Epilepsy	EEG	9,200 / 2,300	1	178	Seizure vs. non-seizure
Sleep-EDF	Sleep EEG (Fpz-Cz, 100 Hz)	25,612 / 8,910	1	3,000	5 sleep stages
Fault Diagnosis (FD-A–D)	Bearing fault diagnosis, 4 operating conditions	8,184 / 2,728	1	5,120	3 (healthy, inner fault, outer fault)
EthanolConcentration	UEA: Alcohol industry	261 / 263	3	1,751	4
FaceDetection	UEA: Face / MEG	5,890 / 3,524	144	62	2
Handwriting	UEA: Handwriting	150 / 850	3	152	26
Heartbeat	UEA: Heart sounds	204 / 205	61	405	2
JapaneseVowels	UEA: Speech	270 / 370	12	29	9
PEMS-SF	UEA: Transportation	267 / 173	963	144	7
SelfRegulationSCP1	UEA: EEG	268 / 293	6	896	2
SelfRegulationSCP2	UEA: EEG	200 / 180	7	1,152	2
SpokenArabicDigits	UEA: Speech	6,599 / 2,199	13	93	10
UWaveGestureLibrary	UEA: Gesture recognition	120 / 320	3	315	8

For Epilepsy, the four non-seizure classes of the original five-class dataset are merged into one, which makes it binary seizure detection. For the UEA datasets, the official train/test splits are used, and validation examples are carved out of the training split. FD-A to FD-D are used for the cross-domain transfer experiments.

Quick start

git clone https://github.com/Sachithx/ConceptTime && cd ConceptTime
pip install -r requirements.txt

python download_data.py                    # HAR + Sleep-EDF test splits and checkpoints
python download_data.py --datasets HAR     # just one

python src/evaluate.py --dataset HAR \
    --artifact checkpoints/HAR/pipeline_best.pt \
    --density  checkpoints/HAR/density_model_best.pt --split test

Each run prints accuracy and macro-F1 at 1%, 5%, and full label budgets. Use --dataset Epilepsy or --dataset SLeep-EDF to evaluate the other headline datasets.

To fetch everything, including the archive, use huggingface_hub:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="sachithabey/ConceptTime",
    repo_type="dataset",
    local_dir="ConceptTime-assets",
    allow_patterns=["archive/datasets/Heartbeat/*", "archive/checkpoints/Heartbeat/*"],  # or omit to get all ~4.3 GB
)

Expected results

Results for ConceptTime-Dynamic, the paper's primary method. The numbers are test accuracy (%) under different label budgets, averaged over 3 seeds (Table 1 of the paper). A single released checkpoint can differ slightly from the 3-seed mean.

Dataset	1% labels	5% labels	Full
UCI-HAR	74.60	79.60	88.91
Epilepsy	97.91	98.01	98.39
Sleep-EDF	70.21	71.92	78.15
UEA (10-dataset average)	58.50	60.10	71.82

ConceptTime-Dynamic is the strongest interpretable method in every column. At 1% labels on Epilepsy, it also beats every black-box baseline (97.91% compared with 95.41% for TSLANet).

Source datasets and licensing

The checkpoints are released under the Apache 2.0 license. The datasets are preprocessed versions of public benchmarks. They are included here only for convenience and reproducibility, and they remain subject to the terms of their original sources. If you use them, please also cite the original authors:

• UCI-HAR: Anguita et al., A Public Domain Dataset for Human Activity Recognition Using Smartphones, ESANN 2013. UCI repository
• Epilepsy: Andrzejak et al., Indications of nonlinear deterministic and finite-dimensional structures in time series of brain electrical activity, Phys. Rev. E, 2001
• Sleep-EDF: Goldberger et al., PhysioBank, PhysioToolkit, and PhysioNet, Circulation, 2000. PhysioNet
• UEA archive: Bagnall et al., The UEA multivariate time series classification archive, 2018, arXiv:1811.00075. timeseriesclassification.com
• Fault Diagnosis: Lessmeier et al., Condition Monitoring of Bearing Damage in Electromechanical Drive Systems, PHM Society European Conference, 2016

Citation

If you find ConceptTime or these assets useful, please cite:

@inproceedings{sachith_concepttime_26,
  title     = {Predictive Surprise as Self-Grounding Concept Bottleneck for Interpretable Time Series},
  author    = {Abeywickrama, Sachith and Eldele, Emadeldeen and Wu, Min and Li, Xiaoli and Yuen, Chau},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2026}
}

Contact

For questions or issues, please open an issue on GitHub.