Research map / Time series / Foundation models for time series
Unsupervised Time-Series Representation Learning with Iterative Bilinear Temporal-Spectral Fusion
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 |
|---|---|---|---|
| Replace segment-level time slicing with instance-level dropout augmentation (rate 0.1) (ablation) | BTSF contrastive pair construction · HAR | Accuracy: 88.3 -> 89.4 (temporal-only ablation) | Helped |
| Add bilinear temporal-spectral fusion instead of temporal-only features (ablation) | BTSF with dropout augmentation · HAR | Accuracy: 89.4 -> 92.4 | Helped |
| Replace bilinear fusion with summation/concatenation of temporal and spectral features (ablation) | BTSF with dropout augmentation · HAR | Accuracy: 89.4 (temporal only) -> 89.8 | Helped |
| Iterate the bilinear fusion (S2T/T2S loops) rather than single pass (ablation) | BTSF · HAR | Accuracy: 92.4 -> 94.6 | Helped |
| Add layer-wise dropout in internal layers (ablation) | BTSF · HAR | Accuracy: 94.6 -> 95.4 (not used in main results) | Helped |
| Use BTSF instead of prior methods on classification | time series classification · HAR, Sleep-EDF, ECG Waveform | Accuracy: HAR 94.63 vs TNC 88.32; exceeds supervised 92.03 | Helped |
| Use BTSF instead of prior methods on forecasting | time series forecasting · ETTh1, ETTh2, ETTm1, Weather (MSE) | MSE: consistently lower MSE than baselines across horizons | Helped |
| Use BTSF instead of prior methods on anomaly detection | anomaly detection · SAaT, WADI, SMD, SMAP, MSL (F1) | F1: e.g., SMD 0.972 vs 0.817 (TNC) | Helped |
| Bilinear fusion with iterative refinement on cross-domain features (temporal-spectral overlap analysis) | BTSF representations · HAR false-prediction overlap | Overlap % of errors by temporal: 33.24% (TNC) -> 96.60% (BTSF) | Helped |
| Dropout rate sensitivity: rate 0.1 vs 0.01 or 0.3 (ablation) | BTSF dropout augmentation · HAR | Accuracy: 0.1 gives 94.63; 0.01 gives 90.29; 0.3 gives 88.07 | Mixed |
| Among 12 augmentation methods, dropout is best (ablation) | BTSF contrastive augmentation · HAR | Accuracy: best average accuracy and variance among augmentations (values in figure only) | Helped |
Closest papers by meaning
- Learning to Embed Time Series Patches Independently (ICLR 2024 · similarity 0.86)
- TimesURL: Self-Supervised Contrastive Learning for Universal Time Series Representation Learning (AAAI 2024 · similarity 0.86)
- TSLANet: Rethinking Transformers for Time Series Representation Learning (ICML 2024 · similarity 0.86)
- T-Rep: Representation Learning for Time Series using Time-Embeddings (ICLR 2024 · similarity 0.85)
- Learning Without Augmenting: Unsupervised Time Series Representation Learning via Frame Projections (NeurIPS 2025 · similarity 0.85)
- Frequency-Masked Embedding Inference: A Non-Contrastive Approach for Time Series Representation Learning (AAAI 2025 · 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.