Research map / Time series / Foundation models for time series
Efficient Source-Free Time-Series Adaptation via Parameter Subspace Disentanglement
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 Tucker-style decomposition of a time-series source model during source preparation, then selectively fine-tunes only the core tensor during source-free target adaptation, reducing fine-tuned parameters and MACs by over 90% while maintaining F1 across SFDA methods.
Problem
Source-free domain adaptation for time-series is parameter- and sample-inefficient, and full-model adaptation is impractical on resource-constrained target devices with few reliable target samples.
Approach
Reparameterize the source model's conv/FC weights with Tucker-style factorization (core tensor plus factor matrices) during source preparation, then fine-tune only the core tensor on the target side; PAC-Bayes analysis argues this constrains parameter distance and thus regularizes adaptation.
Key results
- On SSC, SHOT+SFT at RF=8 reduces fine-tuned parameters by 98.34% (83.17K to 1.38K) and MACs by 93.81% while F1 rises 1.22% on average.
- MAPU+SFT at RF=2 improves average F1 on SSC by 3.54% (63.19 to 65.43) with 74.89% fewer fine-tuned parameters.
- SFT at RF=4 yields 5.02% average F1 gain for AAD on SSC with 93.60% fewer fine-tuned parameters.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add Tucker factorization and core-only SFT to SHOT (RF 8) | SHOT · SSC | F1 score average (%): 64.74 -> 65.53 | Helped |
| Add Tucker factorization and core-only SFT to SHOT (RF 4) | SHOT · SSC | F1 score average (%): 64.74 -> 66.14 | Helped |
| Add Tucker factorization and core-only SFT to NRC (RF 8) | NRC · SSC | F1 score average (%): 62.90 -> 63.10 | Helped |
| Add Tucker factorization and core-only SFT to NRC (RF 4) | NRC · SSC | F1 score average (%): 62.90 -> 64.75 | Helped |
| Add Tucker factorization and core-only SFT to AAD (RF 4) | AAD · SSC | F1 score average (%): 62.10 -> 65.22 | Helped |
| Add Tucker factorization and core-only SFT to MAPU (RF 2) | MAPU · SSC | F1 score average (%): 63.19 -> 65.43 | Helped |
| Increase RF from 4 to 8 in MAPU + SFT (ablation) | MAPU + SFT · SSC | F1 score average (%): 64.94 -> 64.79 | Hurt |
| Fine-tune Batch-Norm parameters only instead of the entire backbone (ablation) | MAPU · SSC | F1 score (%): 66.73 ± 0.85 -> 61.07 ± 1.58 | Hurt |
| Fine-tune core tensor only instead of factor matrices only (ablation) | MAPU + SFT (RF 8) · SSC | F1 score (%): 66.05 ± 0.75 -> 66.09 ± 0.19 | No clear effect |
| Fine-tune factor matrices and core instead of core tensor only (ablation) | MAPU + SFT (RF 8) · SSC | F1 score (%): 66.09 ± 0.19 -> 66.17 ± 0.53 | No clear effect |
| Fine-tune core tensor only instead of factor matrices only (ablation) | MAPU + SFT (RF 4) · SSC | F1 score (%): 67.60 ± 0.07 -> 67.40 ± 0.59 | Hurt |
| Fine-tune core tensor only instead of factor matrices only (ablation) | MAPU + SFT (RF 2) · HHAR | F1 score (%): 80.51 ± 2.41 -> 80.69 ± 2.45 | No clear effect |
Surprising findings
- Fine-tuning only the core tensor outperforms fine-tuning the full decomposed backbone (factor matrices plus core) on several MAPU settings, suggesting the extra capacity hurts.
Limitations
- For RF greater than 8 significant underfitting was observed, limiting the tested range.
- The method is not a new SFDA algorithm; gains depend on the base SFDA method.
Experimental setup
Datasets: SSC MFD HHAR UCIHAR WISDM
Benchmarks: AdaTime SSC AdaTime MFD AdaTime HHAR AdaTime UCIHAR AdaTime WISDM
Metrics: F1 score MACs number of fine-tuned parameters
Baselines: SHOT NRC AAD MAPU LoRA LoKrA
Builds on: SHOT NRC AAD MAPU LoRA
Closest papers by meaning
- Understanding and Improving Source-free Domain Adaptation from a Theoretical Perspective (CVPR 2024 · similarity 0.87)
- Revisiting Source-Free Domain Adaptation: Insights into Representativeness, Generalization, and Variety (CVPR 2025 · similarity 0.87)
- CauDiTS: Causal Disentangled Domain Adaptation of Multivariate Time Series (ICML 2024 · similarity 0.86)
- Diffusion-Driven Progressive Target Manipulation for Source-Free Domain Adaptation (NeurIPS 2025 · similarity 0.86)
- SF(DA)$^2$: Source-free Domain Adaptation Through the Lens of Data Augmentation (ICLR 2024 · similarity 0.86)
- Revisiting the Domain Shift and Sample Uncertainty in Multi-source Active Domain Transfer (CVPR 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.