Learning Compositional Transferability of Time Series for Source-Free Domain Adaptation

TL;DR

This work tackles source-free domain adaptation for time series classification by proposing a compositional reconstruction framework. A frozen pre-trained U-net provides a coarse transfer via a source replay branch, while a warp-based offset compensation branch refines the reconstruction, all balanced by learnable scales and aided by instance-wise test-time rescaling. The approach achieves state-of-the-art MF1 performance on three benchmarks and is validated through extensive ablations and interpretive analyses, highlighting the importance of preserving source priors and enabling fine-grained adaptation. The proposed group-to-instance adaptation offers practical robustness and efficiency for cross-domain time series tasks in settings where source data and target labels are unavailable during training.

Abstract

Domain adaptation is challenging for time series classification due to the highly dynamic nature. This study tackles the most difficult subtask when both target labels and source data are inaccessible, namely, source-free domain adaptation. To reuse the classification backbone pre-trained on source data, time series reconstruction is a sound solution that aligns target and source time series by minimizing the reconstruction errors of both. However, simply fine-tuning the source pre-trained reconstruction model on target data may lose the learnt priori, and it struggles to accommodate domain varying temporal patterns in a single encoder-decoder. Therefore, this paper tries to disentangle the composition of domain transferability by using a compositional architecture for time series reconstruction. Here, the preceding component is a U-net frozen since pre-trained, the output of which during adaptation is the initial reconstruction of a given target time series, acting as a coarse step to prompt the subsequent finer adaptation. The following pipeline for finer adaptation includes two parallel branches: The source replay branch using a residual link to preserve the output of U-net, and the offset compensation branch that applies an additional autoencoder (AE) to further warp U-net's output. By deploying a learnable factor on either branch to scale their composition in the final output of reconstruction, the data transferability is disentangled and the learnt reconstructive capability from source data is retained. During inference, aside from the batch-level optimization in the training, we search at test time stability-aware rescaling of source replay branch to tolerate instance-wise variation. The experimental results show that such compositional architecture of time series reconstruction leads to SOTA performance on 3 widely used benchmarks.

Figures (6)

• Figure 1: Composition of Domain Transferability (The latent features are illustrated in different shapes to denote their belongings to 3 classes, where circle, square, and triangle represent the data instances of class 0, class 1, and class 2 in the MFD data, respectively. Further, blue, red, or green color is assigned to each data instance to denote the source data, the target data before adaptation, and the target data after adaptation, where the black lines indicate the course of migrating the target data from their original positions to the transition points following source replay, marked using black dots, and the subsequent migrating from such transition positions to the destinations after applying additional offset compensation, denoted using red lines). The domain discrepancy is progressively shortened for each class following source replay and offset compensation in order.
• Figure 2: Network architecture.
• Figure 3: Warp block.
• Figure 4: Sensitivity of IA to $\Delta$ and n.
• Figure 5: Visualization of the three classes of MFD data for 0→1 domain transfer.