Research map / Time series / Foundation models for time series
Soft-DTW: a Differentiable Loss Function for Time-Series
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Use a differentiable soft minimum over alignment costs for barycenter optimization | Time-series barycenters, random initialization · UCR 79 datasets, comparison with DBA | datasets with lower DTW loss (%): 97.47% | Helped |
| Use a differentiable soft minimum over alignment costs for barycenter optimization | Time-series barycenters, Euclidean mean initialization · UCR 79 datasets, comparison with DBA | datasets with lower DTW loss (%): 89.87% | Helped |
| Use soft-DTW barycenters in place of DBA centroids | Time-series k-means, random initialization · UCR datasets | datasets with lower DTW loss (%): 77.19% | Helped |
| Use soft-DTW barycenters in place of DBA centroids | Time-series k-means, random initialization · UCR datasets | datasets with lower DTW loss (%): 96.49% | Helped |
| Use soft-DTW barycenters instead of DBA barycenters | Nearest-centroid classification · UCR archive | datasets where accuracy is better or at least as good as DBA (%): 75% | Helped |
| Train the prediction MLP with soft-DTW rather than Euclidean loss | MLP, UCR multistep-ahead prediction · DTW evaluation, random initialization | averaged rank: 3.46 -> 1.87 | Helped |
| Train the prediction MLP with soft-DTW rather than Euclidean loss | MLP, UCR multistep-ahead prediction · Euclidean evaluation, random initialization | averaged rank: 1.05 -> 3.99 | Hurt |
| Reduce smoothing from γ=1 to γ=0.001 (ablation) | Prediction MLP · UCR multistep-ahead prediction, DTW evaluation, random initialization | averaged rank: 3.55 -> 1.87 | Helped |
| Replace Euclidean training loss with soft-DTW (ablation) | Prediction MLP · UCR multistep-ahead prediction, Euclidean evaluation, random initialization | averaged rank: 1.05 -> 2.41 | Hurt |
| Use Euclidean mean initialization rather than random initialization (ablation) | Prediction MLP with soft-DTW loss (γ=0.001) · UCR multistep-ahead prediction, DTW evaluation | averaged rank: 1.87 -> 1.29 | Helped |
Closest papers by meaning
- Fortifying Time Series: DTW-Certified Robust Anomaly Detection (NeurIPS 2025 · similarity 0.85)
- TimePoint: Accelerated Time Series Alignment via Self-Supervised Keypoint and Descriptor Learning (ICML 2025 · similarity 0.85)
- Soft Contrastive Learning for Time Series (ICLR 2024 · similarity 0.84)
- Patch-wise Structural Loss for Time Series Forecasting (ICML 2025 · similarity 0.84)
- Improved Learning via k-DTW: A Novel Dissimilarity Measure for Curves (ICML 2025 · similarity 0.84)
- Learning Pattern-Specific Experts for Time Series Forecasting Under Patch-level Distribution Shift (NeurIPS 2025 · similarity 0.83)
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.