Research map / Time series / Foundation models for time series
TransPL: VQ-Code Transition Matrices for Pseudo-Labeling of Time Series Unsupervised Domain Adaptation
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 TransPL, a pseudo-labeling method for time series UDA that builds class- and channel-wise VQ-code transition matrices from coarse codes, combines them via Bayes' rule with OT-based channel alignment weights, and fine-tunes on pseudo-labeled target data.
Problem
Unsupervised domain adaptation for time series classification suffers from poor pseudo labels because existing pseudo-labeling methods (borrowed from images) ignore temporal transitions and channel-wise shifts between domains.
Approach
Model the source joint P(X,y) with class-wise VQ coarse-code Markov transition matrices, compute channel-wise class-conditional likelihoods, and produce pseudo labels by Bayes' rule weighted by channel alignment scores from optimal transport (earth mover's distance) between source and target channel transition matrices; coarse/fine residual VQ codebooks make the code space tractable.
Key results
- TransPL achieves 69.0% UCIHAR, 64.0% WISDM, 68.4% HHAR, 67.2% PTB accuracy on target test, the best among compared methods and roughly 3.0% accuracy above SHOT on average.
- As a pseudo-labeler, TransPL improves average pseudo-label accuracy by 6.1% over the best baseline; with weak supervision the gain reaches 10.7%.
- Ablation shows channel alignment and weak supervision are complementary; both together give the best results in 7 of 8 cases.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use TransPL instead of the strongest baseline | Time-series UDA · UCIHAR | accuracy: 67.8 -> 69.0 | Helped |
| use TransPL instead of the strongest baseline | Time-series UDA · WISDM | accuracy: 63.7 -> 64.0 | Helped |
| use TransPL instead of the strongest baseline | Time-series UDA · HHAR | accuracy: 64.8 -> 68.4 | Helped |
| use TransPL instead of the strongest baseline | Time-series UDA · PTB | accuracy: 62.4 -> 67.2 | Helped |
| add channel alignment (CA) (ablation) | TransPL · UCIHAR | accuracy: 68.6 -> 71.2 | Helped |
| add channel alignment (CA) (ablation) | TransPL · WISDM | accuracy: 69.2 -> 71.3 | Helped |
| add channel alignment (CA) (ablation) | TransPL · HHAR | accuracy: 62.3 -> 70.4 | Helped |
| add channel alignment (CA) (ablation) | TransPL · PTB | accuracy: 72.2 -> 72.4 | Helped |
| add weak-supervision label prior (WS) (ablation) | TransPL · UCIHAR | accuracy: 69.0 -> 71.2 | Helped |
| add weak-supervision label prior (WS) (ablation) | TransPL · WISDM | accuracy: 64.0 -> 71.3 | Helped |
| add weak-supervision label prior (WS) (ablation) | TransPL · HHAR | accuracy: 68.4 -> 70.4 | Helped |
| add weak-supervision label prior (WS) (ablation) | TransPL · PTB | accuracy: 67.2 -> 72.4 | Helped |
Surprising findings
- CoDATS degrades with weak supervision on UCIHAR, HHAR and PTB, suggesting label-distribution supervision can introduce prediction bias.
- Discriminative sequence models (1D-CNN, LSTM, GRU) give much worse pseudo-labels than the transition-matrix approach.
Limitations
- Channel weighting penalizes shifted channels, which can hurt when class-relevant information lies in heavily shifted channels.
- Coarse/fine hierarchy is encouraged empirically but not enforced by regularization.
- Results depend on tuning of temperature tau and RBF bandwidth sigma per task.
Experimental setup
Datasets: UCIHAR WISDM HHAR PTB
Benchmarks: UCIHAR WISDM HHAR PTB
Metrics: accuracy macro F1
Baselines: DeepCoral MMDA CoDATS SASA RAINCOAT SHOT T2PL SoftMax
Builds on: VQVAE SHOT CoDATS AdaTime
Closest papers by meaning
- Pseudo-Labeling for Unsupervised Domain Adaptation with Kernel GLMs (NeurIPS 2026 · similarity 0.86)
- Enhancing Multivariate Time-Series Domain Adaptation via Contrastive Frequency Graph Discovery and Language-Guided Adversary Alignment (AAAI 2025 · similarity 0.85)
- Test-Time Domain Adaptation by Learning Domain-Aware Batch Normalization (AAAI 2024 · similarity 0.85)
- Pseudo-Calibration: Improving Predictive Uncertainty Estimation in Unsupervised Domain Adaptation (ICML 2024 · similarity 0.85)
- Everything to the Synthetic: Diffusion-driven Test-time Adaptation via Synthetic-Domain Alignment (CVPR 2025 · similarity 0.85)
- Domain-Specific Block Selection and Paired-View Pseudo-Labeling for Online Test-Time Adaptation (CVPR 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.