Research map / Time series / Foundation models for time series
Self-Supervised Framework Based on Subject-Wise Clustering for Human Subject Time Series Data
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 SA-TSC, a self-supervised framework for human subject time series that trains with subject-wise batches, graph-based spatial encoders, and spectral clustering over subject graphs, evaluated in limited subject-level label settings.
Problem
Self-supervised time series learning for human subject data ignores subject-level annotation and large inter-subject variation, which matter in clinical settings with scarce labels.
Approach
Subject-wise (rather than instance-wise) batching with subject graphs built from Gaussian-kernel similarity and Gumbel-Softmax edge sampling, combined with spectral clustering loss and a graph-based agreement loss on positive/negative pairs from edge probabilities.
Key results
- In the 2-subject fine-tuning setting, SA-TSC reaches 93.11% ACC / 93.18% F1 on HAR versus 83.85% / 82.72% for CA-TCC.
- On SleepEDF with 1 subject labeled, SA-TSC achieves 73.11% ACC / 65.43% F1, above CA-TCC's 70.51% / 63.20%.
- Ablation shows the full model (Gaussian & Gumbel graph construction) reaches 75.24% ACC on SleepEDF versus 73.21% without the clustering loss-related components in the table.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use subject-wise clustering and subject-level fine-tuning | SA-TSC · HAR 1sub | ACC: 87.14 ± 0.63 -> 91.32 ± 0.90 | Helped |
| Use subject-wise clustering and subject-level fine-tuning | SA-TSC · ISRUC 1sub | ACC: 69.27 ± 0.74 -> 69.63 ± 0.36 | Helped |
| Use subject-wise clustering and subject-level fine-tuning | SA-TSC · SleepEDF 1sub | ACC: 70.90 ± 0.52 -> 73.11 ± 0.41 | Helped |
| Use subject-wise clustering and subject-level fine-tuning | SA-TSC · HAR 2sub | ACC: 83.85 ± 1.66 -> 93.11 ± 1.82 | Helped |
| Use subject-wise clustering and subject-level fine-tuning | SA-TSC · ISRUC 2sub | ACC: 73.03 ± 0.96 -> 72.69 ± 0.41 | Hurt |
| Use subject-wise clustering and subject-level fine-tuning | SA-TSC · SleepEDF 2sub | ACC: 74.92 ± 0.47 -> 76.01 ± 0.01 | Helped |
| Add agreement loss LA (ablation) | SA-TSC spatial representation learning · SleepEDF 2sub | ACC: 69.68 -> 73.21 | Helped |
| Add subject graph clustering loss LSP (ablation) | SA-TSC · SleepEDF 2sub | ACC: 73.21 -> 75.24 | Helped |
| Use Gaussian and Gumbel graph construction rather than Gaussian alone (ablation) | SA-TSC · SleepEDF 2sub | ACC: 70.11 -> 75.24 | Helped |
| Use Gaussian and Gumbel graph construction rather than Gumbel alone (ablation) | SA-TSC · SleepEDF 2sub | ACC: 71.32 -> 75.24 | Helped |
| Add stop-gradient operator so cluster assignment is optimized based on fixed spatial representation (ablation) | SA-TSC · HAR, ISRUC, and SleepEDF 2sub | ACC: 93.73 -> 96.24; 66.32 -> 72.69; 72.75 -> 75.91 | Helped |
| Add stop-gradient operator so cluster assignment is optimized based on fixed spatial representation (ablation) | SA-TSC · HAR, ISRUC, and SleepEDF 2sub | F1: 94.06 -> 96.37; 63.87 -> 69.54; 65.44 -> 70.77 | Helped |
Surprising findings
- On ISRUC with 2 subjects, the baseline CA-TCC slightly outperforms SA-TSC in F1 (69.94 vs 69.54), though SA-TSC has lower variance.
Limitations
- Graph sparsity controlled by Gaussian kernel length scale sigma is sensitive and needs careful tuning.
- Computation cost grows as the number of graph nodes increases.
- Evaluated on only three datasets with few subjects.
Experimental setup
Datasets: UCI-HAR ISRUC SleepEDF
Benchmarks: UCI-HAR ISRUC SleepEDF
Metrics: accuracy F1-score
Baselines: Random Initialization CoSleep TS-TCC CA-TCC
Builds on: TS-TCC CA-TCC Barlow Twins DiffPool-style spectral clustering loss
Closest papers by meaning
- Sequential Multi-Dimensional Self-Supervised Learning for Clinical Time Series (ICML 2023 · similarity 0.86)
- Integrating Sequence and Image Modeling in Irregular Medical Time Series Through Self-Supervised Learning (AAAI 2025 · similarity 0.83)
- Medformer: A Multi-Granularity Patching Transformer for Medical Time-Series Classification (NeurIPS 2024 · similarity 0.82)
- Not All Data are Good Labels: On the Self-supervised Labeling for Time Series Forecasting (NeurIPS 2025 · similarity 0.82)
- Shape analysis for time series (NeurIPS 2024 · similarity 0.81)
- Clustering Properties of Self-Supervised Learning (ICML 2025 · similarity 0.81)
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.