Time Series Domain Adaptation via Multiscale Feature Extraction and Mix-up

Conducting Unsupervised Domain Adaptation (UDA) on Time Series (TS) data is an important and challenging task. The majority of works in this area focus on extracting domain-invariant representations from the data, often achieved by employing constraints such as Maximum Mean Discrepancy (MMD) or introducing an adversarial opponent. However, this extraction process is non-trivial due to the non-linear and non-stationary relationships existing between different timestamps. Moreover, when reasoning on higher level features which consists multiple timestamps, the presence of noise and variations in timescales across diverse classes and domains further complicates the extraction of stable and reliable features. Therefore, the core challenge lies in overcoming these complexities to adaptively extract effective features with multiscale information, which is capable of representing the non-trivial relationships within temporal data.In this paper, we propose a novel framework named MSEM designed to extracting domain-invariant representations through scale-awareness feature extraction and fusion. Initially, we use multiscale group convolution on TS input data to obtain semantic information from various time length with diverse receptive fields. Subsequently, a multiscale mix-up module seamlessly integrates TS data from both domains at the feature level, with respect to its information in both short and long terms. With the help of mix-up loss, the entire framework enhances the domain adversarial network’s ability to further manipulate the learned representation, ensuring its domain-invariance while preserving optimal scale information for classification. This results in a more robust fused feature, contributing to improved performance. Extensive experimental studies not only demonstrate its state-of-the-art performance across five real-world datasets, but also verify the effectiveness of our method.