Research map / Time series / Foundation models for time series
Learning Without Augmenting: Unsupervised Time Series Representation Learning via Frame Projections
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 an augmentation-free SSL method for time series that contrasts a time-domain view with orthonormal (Fourier) and overcomplete frame (Gabor wavelet) views, plus lightweight latent mappers for linear probing.
Problem
Self-supervised representation learning for temporal signals where handcrafted augmentations are hard to design and impose invariances that can hurt downstream tasks.
Approach
Replace augmentations with fixed unitary (Fourier) and redundant frame (Gabor wavelet) projections as views, apply NT-Xent instance discrimination across time, Fourier and wavelet encoders, and train non-linear mappers that predict other-domain latents from the time-domain latent so that the concatenated features exploit distinct manifold geometries.
Key results
- On HHAR, the method reaches 70.67% accuracy versus 68.56% for the best baseline (TS-TCC); on USC it reaches 52.21% versus 35.40% for TS2Vec.
- On Sleep-EDF, accuracy is 77.30% versus 74.69% for SimMTM, with the paper claiming gains of up to 15-20% over existing SSL methods.
- Ablations show removing the Fourier or wavelet view or the mappers generally degrades performance, though the wavelet-only variant sometimes wins.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use frame-projection SSL rather than the strongest listed SSL baseline | Ours · IEEE SPC12 | MAE: 9.80 $\pm$0.49 -> 8.84 $\pm$0.50 | Helped |
| Use frame-projection SSL rather than the strongest listed SSL baseline | Ours · IEEE SPC22 | MAE: 14.12 $\pm$0.36 -> 14.06 $\pm$1.09 | Helped |
| Use frame-projection SSL rather than the strongest listed SSL baseline | Ours · DaLiA | MAE: 10.23 $\pm$0.01 -> 9.13 $\pm$0.20 | Helped |
| Use frame-projection SSL rather than the strongest listed SSL baseline | Ours · HHAR | accuracy: 68.56 $\pm$1.19 -> 70.67 $\pm$0.06 | Helped |
| Use frame-projection SSL rather than the strongest listed SSL baseline | Ours · USC | accuracy: 35.40 $\pm$0.96 -> 52.21 $\pm$1.09 | Helped |
| Use frame-projection SSL rather than the strongest listed SSL baseline | Ours · Clemson | MAE: 2.70 $\pm$0.06 -> 2.50 $\pm$0.13 | Helped |
| Use frame-projection SSL rather than the strongest listed SSL baseline | Ours · Chapman | accuracy: 85.84 $\pm$0.39 -> 87.21 $\pm$0.80 | Helped |
| Use frame-projection SSL rather than the strongest listed SSL baseline | Ours · CPSC | accuracy: 51.90 $\pm$0.30 -> 52.10 $\pm$0.90 | Helped |
| Use frame-projection SSL rather than the strongest listed SSL baseline | Ours · Sleep-EDF | accuracy: 74.69 $\pm$1.84 -> 77.30 $\pm$1.04 | Helped |
| add the Fourier orthonormal-basis transformation (ablation) | Ours · IEEE SPC12 | MAE: 11.23 -> 8.84 | Helped |
| add the Gabor overcomplete-frame transformation (ablation) | Ours · IEEE SPC12 | MAE: 9.80 -> 8.84 | Helped |
| add latent representation mappers (ablation) | Ours · IEEE SPC12 | MAE: 9.62 -> 8.84 | Helped |
Surprising findings
- Using only the Gabor wavelet view sometimes outperforms the full Fourier+wavelet combination (e.g., CPSC and Sleep ablation without OB).
- Adding a third augmentation-based view helps prior methods but they still trail the proposed method by about 10%.
Limitations
- No formal theoretical explanation for the observed gains; the implicit-bias hypothesis is unverified.
- Experiments focus on classification and regression tasks, not forecasting.
- Wavelet transform computation is expensive, mitigated by caching.
Experimental setup
Datasets: IEEE SPC12 IEEE SPC22 DaLiA HHAR USC Clemson CPSC2018 Chapman Sleep-EDF
Benchmarks: IEEE SPC12 IEEE SPC22 DaLiA HHAR USC Clemson CPSC2018 Chapman Sleep-EDF
Metrics: MAE RMSE Spearman rho accuracy weighted F1 F1 MAPE AUC Cohen's kappa
Baselines: SimCLR BYOL VICReg Barlow Twins CLIP TS-TCC TF-C SimMTM TS2Vec
Builds on: SimCLR TF-C NT-Xent
Closest papers by meaning
- Joint‑Embedding vs Reconstruction: Provable Benefits of Latent Space Prediction for Self‑Supervised Learning (NeurIPS 2025 · similarity 0.87)
- URLOST: Unsupervised Representation Learning without Stationarity or Topology (ICLR 2025 · similarity 0.87)
- Self-supervised Representation Learning from Random Data Projectors (ICLR 2024 · similarity 0.87)
- Frequency-Masked Embedding Inference: A Non-Contrastive Approach for Time Series Representation Learning (AAAI 2025 · similarity 0.87)
- Integrating Sequence and Image Modeling in Irregular Medical Time Series Through Self-Supervised Learning (AAAI 2025 · similarity 0.86)
- TimesURL: Self-Supervised Contrastive Learning for Universal Time Series Representation Learning (AAAI 2024 · similarity 0.86)
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.