Transferability-Driven Variable Recalibration for Multivariate Time Series Domain Adaptation

Kexuan Zhou, Jiajing Geng, Boliang Hao, Bailing Zhang

Abstract: Unsupervised Domain Adaptation (UDA) seeks to transfer knowledge from a labeled source domain to an unlabeled target domain. However, when applied to multivariate time series (MTS) data, UDA faces unique challenges arising from the inherent complexity of sensor-generated signals. While most existing UDA approaches rely on feature extractors to capture global representations, they often overlook the heterogeneous distribution shifts across individual variables—an issue primarily caused by variations in sensor properties and recording conditions. In this paper, we systematically investigate variable-level transferability in multivariate time series (MTS) data and report two key findings: (1) Transferability varies significantly across different variables, and (2) aligning marginal distributions plays a more crucial role in reducing domain discrepancy than adapting conditional distributions. Motivated by these insights, we propose a novel method called Transferability-Driven Variable Recalibration (TDVR), which comprises three core components: (1) Variable-Specific Marginal Distribution Modeling (VSMDM): Each variable is individually processed using a dedicated 1D convolutional neural network (1D-CNN) to extract domain-invariant marginal features; (2) Quantitative Transferability Alignment (QTA): We leverage Maximum Mean Discrepancy (MMD) to measure variable-wise transferability and dynamically recalibrate their distributions accordingly; (3) Prototype-Guided Adaptive Fusion (PGAF): During inference, predictions in the target domain are refined by aligning them with class-specific prototypes derived from the source domain in the latent space. Extensive experiments on diverse time series benchmarks demonstrate that TDVR consistently outperforms existing methods, achieving new state-of-the-art performance.

External IDs:dblp:conf/ecai/ZhouGHZ25