Research map / Time series / Foundation models for time series
Divide and Contrast: Learning Robust Temporal Features without Augmentation
Abdul-Kazeem Shamba, Kerstin Bach, Gavin Taylor
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add 50% overlap between sub-blocks (ablation) | Di-COT · Large datasets | average accuracy (%): 81.22 -> 83.07 | Improved |
| add 50% overlap between sub-blocks (ablation) | Di-COT · UCR | average accuracy (%): 81.12 -> 81.33 | Improved |
| add 50% overlap between sub-blocks (ablation) | Di-COT · UEA | average accuracy (%): 70.13 -> 71.24 | Improved |
| add temperature scaling (tau=0.07) (ablation) | Di-COT · Large, UEA, UCR | average accuracy (%): 82.47 -> 83.07 (Large); no change on UCR (81.33) | Mixed |
| sample sub-block splits uniformly per iteration instead of a fixed global split (ablation) | Di-COT · UEA | average accuracy (%): 69.69 -> 71.24 | Improved |
| sample sub-block splits uniformly per iteration instead of a fixed global split (ablation) | Di-COT · UCR | average accuracy (%): 81.32 -> 81.33 | No clear effect |
| use preceding sub-block as positive instead of shuffled sub-blocks (ablation) | Di-COT · Large datasets | average accuracy (%): 81.80 -> 83.07 | Improved |
| use preceding instead of next sub-block contrast (ablation) | Di-COT · Large datasets | average accuracy (%): 82.22 -> 83.07 | Improved |
| use InceptionTime backbone instead of FCN or ResNet (ablation) | Di-COT · Large datasets | average accuracy (%): 80.11 (ResNet) -> 83.07 | Improved |
| add non-linear projection layer (ablation) | Di-COT · Large datasets | average accuracy (%): 83.07 -> 81.85 | Hurt |
| contrast sub-blocks instead of every timestep (Di-COT vs CaTT) | temporal contrastive learning · UCR (124 datasets) | average accuracy (%): 77.42 -> 81.33 | Improved |
| contrast sub-blocks instead of every timestep (Di-COT vs CaTT) | temporal contrastive learning · six large-scale datasets, training time | training time (hours): 3.47 -> 2.88 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- TS2Vec — TS2Vec: Towards Universal Representation of Time Series (AAAI 2022) · better on 2 of 2 shared results
- CoST — CoST: Contrastive Learning of Disentangled Seasonal-Trend Representations for Time Series Forecasting (ICLR 2022) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- TimesURL: Self-Supervised Contrastive Learning for Universal Time Series Representation Learning (AAAI 2024 · similarity 0.88)
- Soft Contrastive Learning for Time Series (ICLR 2024 · similarity 0.87)
- Frequency-Masked Embedding Inference: A Non-Contrastive Approach for Time Series Representation Learning (AAAI 2025 · similarity 0.86)
- TRACE: Contrastive learning for multi-trial time series data in neuroscience (NeurIPS 2025 · similarity 0.86)
- Learning Without Augmenting: Unsupervised Time Series Representation Learning via Frame Projections (NeurIPS 2025 · similarity 0.86)
- Parametric Augmentation for Time Series Contrastive Learning (ICLR 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
