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

TL;DR

TransPL tackles time-series unsupervised domain adaptation by explicitly modeling temporal transitions and channel shifts through coarse/fine vector-quantized code transition matrices. It constructs class-wise and channel-wise TMs from source data and uses Bayes' rule, weighted by channel alignment scores from optimal transport, to generate pseudo-labels for the target domain. The target model is then fine-tuned with a pseudo-labeled target set, using a joint loss that includes VQ reconstruction and classification terms, with optional weak supervision via a known target label distribution. Across four benchmarks, TransPL achieves consistent improvements over state-of-the-art pseudo-labeling methods and offers interpretable insights into the domain shift dynamics via the learned transition matrices.

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 $P(\mathbf{X}, y)$ 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.

Figures (6)

• Figure 1: Overall scheme of TransPL. (A) Source Training: We first train the whole model architecture, i.e., encoder, decoder, two VQ codebooks (coarse and fine codebook), and classifier, using the labeled source domain data. A [CLS] token is appended to the input patches and is used as the input to the classifier. (B) Coarse Code Transition Matrices (TM): The trained encoder and coarse codebook infer coarse codes from both target (unlabeled) and source (labeled) domains. These codes serve as states for constructing class-wise TM (from source) and channel-wise TM (from both domains). (C) Pseudo Label Construction: For unlabeled target data, we compute class-conditional likelihoods per channel using class-wise TM to obtain channel-wise class labels. These are weighted by the similarity between source and target channel-wise TMs, then averaged to generate final pseudo labels.
• Figure 2: A. Average PE: The average permutation entropy (PE) score for the reconstructed coarse and fine VQ codes. PE is a measure of temporal complexity bandt2002permutation, where a lower PE indicates a simpler pattern. Across all time series benchmarks (i.e. UCIHAR, HHAR, WISDM), the coarse code exhibits lower PE compared to the fine code. B. Visualization: The reconstructed coarse and fine VQ codes shows that the coarse code has simplistic patterns (e.g. upward, downward trends), while the fine code has more complex patterns. Full results in \ref{['appendix:coarse_and_fine']}.
• Figure 3: Class conditional likelihood visualization in UCI. Samples from the source are used to construct the class-wise transition matrices (TMs). Then, these TMs are used to calculate the class-conditional likelihoods of the unlabeled target sequence.
• Figure 4: Channel alignment (CA) analysis. We visualized the latent representation of each channels's [CLS] token and observe that the degree of shift differs between channels. Specifically, channel 3 shows higher degrees of shift between source and target compared to channel 1 and 2. We show that using prototype distances (as in clustering based approaches) lead to inaccurate distance measurement, while our earth mover's distance from the CA module provides well calibrated distance between channels.
• Figure 5: Visualization of the full coarse and fine codes in HHAR task. We visualized the trained coarse and fine codes from the HHAR task $2\rightarrow11$. The top row contains the 8 coarse codes, and the rest are 64 fine codes. We observe that the fine codes are capturing more fine grained detailed compared to the coarse codes.