GLA-DA: Global-Local Alignment Domain Adaptation for Multivariate Time Series

TL;DR

A Global-Local Alignment Domain Adaptation (GLA-DA) method for multivariate time series data, which aims to preserve differences among data with distinct labels by aligning the samples with the same class labels together, achieving Local Class Alignment (LCA).

Abstract

Unlike images and natural language tokens, time series data is highly semantically sparse, resulting in labor-intensive label annotations. Unsupervised and Semi-supervised Domain Adaptation (UDA and SSDA) have demonstrated efficiency in addressing this issue by utilizing pre-labeled source data to train on unlabeled or partially labeled target data. However, in domain adaptation methods designed for downstream classification tasks, directly adapting labeled source samples with unlabelled target samples often results in similar distributions across various classes, thereby compromising the performance of the target classification task. To tackle this challenge, we proposed a Global-Local Alignment Domain Adaptation (GLA-DA) method for multivariate time series data. Data from two domains were initially encoded to align in an intermediate feature space adversarially, achieving Global Feature Alignment (GFA). Subsequently, GLA-DA leveraged the consistency between similarity-based and deep learning-based models to assign pseudo labels to unlabeled target data. This process aims to preserve differences among data with distinct labels by aligning the samples with the same class labels together, achieving Local Class Alignment (LCA). We implemented GLA-DA in both UDA and SSDA scenarios, showcasing its superiority over state-of-the-art methods through extensive experiments on various public datasets. Ablation experiments underscored the significance of key components within GLA-DA.

Figures (4)

• Figure 1: Possible data distributions after global feature alignment.
• Figure 2: An overview of our model GLA-DA for Multivariate Time Series. The dashed lines indicate the parameters of that model are fixed. First we trained a source model with labeled source data. Second, we assigned initial pseudo labels for unlabeled target dataset just in unsupervised scenario based on the threshold output by the source model. Third, we assigned pseudo labels to the unlabeled target samples left using AM. Forth, we performed global data alignment using adversarial learning and local alignment by optimizing center-loss. Last we trained a shared classifier.
• Figure 3: Architecture of 1D-CNN encoder used in our proposed GLA-DA.
• Figure 4: Effect of local class alignment compared to Source Only. Red and its different color depth indicates the source domain and blue and its different color depth indicates the target domain.