Research map / Time series / Foundation models for time series
TimePoint: Accelerated Time Series Alignment via Self-Supervised Keypoint and Descriptor Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes TimePoint, a self-supervised 1D keypoint detector and descriptor network trained on synthetic signals warped by CPAB diffeomorphisms, so DTW can align sparse keypoints faster and often more accurately.
Problem
Time series alignment with Dynamic Time Warping is quadratic in sequence length and sensitive to noise, and there is no keypoint/descriptor learning framework for 1D signals comparable to SuperPoint in images.
Approach
Adapts SuperPoint-style self-supervised keypoint and descriptor learning to 1D signals: synthetic signals with known keypoints are warped with CPAB diffeomorphisms to create ground-truth correspondences, and a fully convolutional wavelet-convolution (WTConv) encoder with keypoint and descriptor decoders yields sparse features on which cosine-cost DTW runs.
Key results
- On 102 UCR datasets, DTW with TimePoint keypoints at 20% achieves 0.721 vs 0.710-0.732 for full DTW on the synthetic-only model, and with fine-tuning 0.790 at 20% vs 0.732 for full DTW; runtime drops from 192 to about 3 GPU hours (~x65 speedup).
- TP+DTW achieves the highest average rank in the critical difference diagram at 100% and 20% keypoint usage, with synthetic-only training.
- Fine-tuning on real UCR data improves accuracy by roughly 7-8% across keypoint ratios without changing runtime.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Fine-tune TimePoint on real data | TimePoint + DTW · 102 UCR datasets | 1-NN accuracy: 0.721 -> 0.790 | Helped |
| Use 20% rather than 100% of TimePoint keypoints | TimePoint + DTW · 102 UCR datasets | GPU runtime (hours): 193 -> 2.88 | Helped |
| Use TimePoint descriptors at 20% keypoints instead of raw DTW | DTW · 102 UCR datasets | 1-NN accuracy: 0.706 -> 0.721 | Helped |
| Use TimePoint descriptors at 20% keypoints instead of raw DTW | DTW · 102 UCR datasets | GPU runtime (hours): 192 -> 2.88 | Helped |
| Use cosine similarity instead of Euclidean distance (ablation) | TimePoint with Dense Conv encoder, 100% length · 27 datasets | 1-NN accuracy: 0.82 -> 0.835 | Helped |
| Use cosine similarity instead of Euclidean distance (ablation) | TimePoint with WTConv encoder, 100% length · 27 datasets | 1-NN accuracy: 0.815 -> 0.869 | Helped |
| Use WTConv rather than Dense Conv encoder with cosine similarity (ablation) | TimePoint, 100% length · 27 datasets | 1-NN accuracy: 0.835 -> 0.869 | Helped |
| Add descriptors rather than use raw signal restricted to keypoints (ablation) | TimePoint + DTW, 20% length · 27 datasets | 1-NN accuracy: 0.792 -> 0.865 | Helped |
| Add non-maximum suppression (ablation) | TimePoint with WTConv and cosine similarity, 20% length · 27 datasets | 1-NN accuracy: 0.841 -> 0.865 | Helped |
| Use TimePoint at 20% rather than 10% keypoints | TimePoint · 30 UCR datasets, no perturbation | 1-NN accuracy: 0.867 -> 0.881 | Helped |
| Use TimePoint at 20% rather than 10% keypoints | TimePoint · 30 UCR datasets, jitter σ = 0.5 | 1-NN accuracy: 0.760 -> 0.791 | Helped |
Surprising findings
- Sparse DTW on 10-20% of keypoints often yields higher 1-NN accuracy than full-length DTW.
- Using TimePoint only as subsampling (no descriptors) gives lower accuracy than with descriptors, though still below the full-length cosine variant.
Limitations
- Performance may be suboptimal when real signals deviate substantially from the synthetic distribution, requiring fine-tuning.
- Temporal distortions beyond the CPAB prior may require adjustments.
- Fixed downsampling factor of 8 may over-compress short sequences; only univariate signals are handled.
Experimental setup
Datasets: SynthAlign UCR Time Series Archive
Benchmarks: UCR Time Series Archive
Metrics: 1-NN classification accuracy GPU runtime (hours) average rank (critical difference diagram)
Baselines: DTW SoftDTW DTW-GI ShapeDTW Euclidean
Builds on: SuperPoint CPAB transformations WTConv Dynamic Time Warping DTAN
Closest papers by meaning
- TimeDART: A Diffusion Autoregressive Transformer for Self-Supervised Time Series Representation (ICML 2025 · similarity 0.85)
- Time Series Representations with Hard-Coded Invariances (ICML 2025 · similarity 0.85)
- TimeTok: Granularity-Controllable Time-Series Generation via Hierarchical Tokenization (NeurIPS 2026 · similarity 0.85)
- TimeBase: The Power of Minimalism in Efficient Long-term Time Series Forecasting (ICML 2025 · similarity 0.85)
- TimeSiam: A Pre-Training Framework for Siamese Time-Series Modeling (ICML 2024 · similarity 0.84)
- T-Rep: Representation Learning for Time Series using Time-Embeddings (ICLR 2024 · similarity 0.84)
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.