This study proposes a three-step domain adaptation method to address the diminished performance observed in the time series data classification. This decline stems from (i) dissimilarities in data distribution between the source and target domains and...
This study proposes a three-step domain adaptation method to address the diminished performance observed in the time series data classification. This decline stems from (i) dissimilarities in data distribution between the source and target domains and (ii) the absence of labeled data in the target domain. The efficacy of the proposed method is substantiated through its application to a tailored human activity recognition (HAR) dataset. Our approach commenced by training the time series classification models employing adversarial backpropagation. Subsequently, a pseudo-labeling technique was applied to the unlabeled target domain data. This technique generated pseudo-labels through an ensemble voting scheme using the domain-adapted classification models. Finally, a specially chosen domain-adapted model underwent fine-tuning, incorporating both the labeled source and pseudo-labeled target domain data. The outcomes of our study indicate that integrating domain adaptation with adversarial learning and pseudo-labeling yields a notable enhancement in performance in the target domain.