TransPL: VQ-Code Transition Matrices for Pseudo-Labeling of Time Series Unsupervised Domain Adaptation

Jaeho Kim, Seulki Lee

Introduction

Unsupervised Domain Adaptation (UDA) in deep learning aims to transfer knowledge from a labeled "source" domain to an unlabeled "target" domain. For time series data, such as wearable sensor readings or ECGs, UDA faces two unique hurdles: temporal dynamics and channel-wise shifts. Traditional pseudo-labeling strategies, often borrowed from computer vision, typically ignore sequential dependencies and treat domain shifts as a "black box." This can lead to noisy pseudo-labels, especially when multi-sensor setups experience hardware differences or placement shifts.

To address these limitations, researchers have introduced TransPL (Transition-based Pseudo-Labeling), a novel UDA framework that replaces black-box pseudo-labeling with an explainable approach based on temporal transitions. By modeling time series patches as discrete codes and tracking their transitions via Markov chains, TransPL explicitly accounts for temporal dynamics and channel-wise sensor shifts.

Methodology: The TransPL Framework

TransPL operates through a three-stage pipeline designed to capture and transfer temporal transition patterns across domains.

The overall TransPL architecture, showing source training, transition matrix construction, and pseudo-label generation.

1. Coarse and Fine Vector Quantization (VQ)

Instead of processing raw continuous signals, TransPL breaks time series into patches and maps them to discrete "codes" using a VQ-VAE-inspired architecture. This mapping is hierarchical:

  • Coarse Codebook: Captures essential, low-frequency "motifs" or global trends of the signal.
  • Fine Codebook: Captures high-frequency residuals and fine-grained details.

2. Modeling Transitions via Markov Chains

The core innovation of TransPL is treating the sequence of coarse codes as a Markov chain. The model builds Transition Matrices (TMs) that represent the probability of one state (code) transitioning to another:

  • Class-wise TMs: Learned from the source domain to define the characteristic temporal transitions of each class.
  • Channel-wise TMs: Constructed for both source and target domains to identify how specific sensors (channels) shift across domains.

3. Explainable Pseudo-Labeling via Bayes' Rule

To label target data, TransPL employs Bayes' rule, calculating the likelihood of a target sequence belonging to a specific class based on the source's transition patterns. To handle sensor shifts, the framework uses Optimal Transport to compare source and target channel-wise TMs. If a channel's transition patterns have shifted significantly, the model assigns it a lower "Channel Alignment Score," prioritizing reliable sensors for the final prediction.

Key Insights and Results

TransPL was evaluated on four major time series benchmarks: UCIHAR, WISDM, and HHAR (human activity recognition), as well as PTB (electrocardiogram classification).

Performance Gains

TransPL consistently outperformed state-of-the-art baselines like SHOT and CoDATS. It achieved an average 6.1% improvement in pseudo-label accuracy and a 4.9% improvement in Macro-F1 scores across the benchmarks. Furthermore, the framework naturally accommodates weak supervision (prior knowledge of the target domain's label distribution), which widens its performance gap over competitors.

Coarse Codes for Stable Transitions

By analyzing Permutation Entropy (PE), the authors demonstrated that coarse codes represent simpler, more stable global trends. This makes them significantly more effective for modeling robust temporal transitions than high-entropy fine codes.

Selective Channel Adaptation

Not all sensors are equally affected by domain shifts. By weighting channels based on their transition alignment, TransPL avoids being misled by a single noisy or shifted sensor, such as a loosely fitted smartwatch gyroscope.

Explainability

Unlike standard neural networks, TransPL provides interpretable insights into the domain adaptation process. Researchers can inspect the transition matrices to understand why a model classified a segment a certain way.

Visualization of class-conditional likelihoods, showing how the model identifies the correct class based on transition patterns.

Limitations and Future Work

While highly effective, TransPL has two primary areas for future improvement:

  1. Information vs. Shift Trade-off: Currently, TransPL down-weights channels that shift heavily. However, a heavily shifted channel might still contain critical classification information. Future iterations could integrate "channel importance" metrics to balance this trade-off.
  2. Mathematical Regularization: The coarse/fine hierarchy works well empirically but lacks a strict mathematical constraint to guarantee that coarse codes always capture the most general patterns. Adding specific regularization terms could make this hierarchical representation more robust.

Summary updated

Abstract

Unsupervised domain adaptation (UDA) for time series data remains a critical challenge in deep learning, with traditional pseudo-labeling strategies failing to capture temporal patterns and channel-wise shifts between domains, producing sub-optimal pseudo labels. As such, we introduce TransPL, a novel approach that addresses these limitations by modeling the joint distribution $P(X,y)$ of the source domain through code transition matrices, where the codes are derived from vector quantization (VQ) of time series patches. Our method constructs class- and channel-wise code transition matrices from the source domain and employs Bayes' rule for target domain adaptation, generating pseudo-labels based on channel-wise weighted class-conditional likelihoods. TransPL offers three key advantages: explicit modeling of temporal transitions and channel-wise shifts between different domains, versatility towards different UDA scenarios (e.g., weakly-supervised UDA), and explainable pseudo-label generation. We validate TransPL's effectiveness through extensive analysis on four time series UDA benchmarks and confirm that it consistently outperforms state-of-the-art pseudo-labeling methods by a strong margin (6.1\% accuracy improvement, 4.9\% F1 improvement), while providing interpretable insights into the domain adaptation process through its learned code transition matrices.
BibTeX
@inproceedings{kim2025transpl,
  author = {Jaeho Kim and Seulki Lee},
  title = {{TransPL: VQ-Code Transition Matrices for Pseudo-Labeling of Time Series Unsupervised Domain Adaptation}},
  booktitle = {International Conference on Machine Learning},
  year = {2025},
  url = {https://openreview.net/forum?id=0AUltNAxjP},
}
Figure