Relation-preserving masked modeling for semi-supervised time-series classification

In this study, we address the challenge of label sparsity in time-series classification using semi-supervised learning that effectively leverages numerous unlabeled instances. Our approach introduces a pioneering framework for semi-supervised time-series classification based on masked time-series modeling, a recent advancement in self-supervised learning that can effectively capture intricate temporal structures in time series. The proposed method first extracts the intrinsic semantic information from unlabeled instances by considering diverse temporal resolutions and using various masking ratios during model training. Subsequently, we combine the semantic information captured from unlabeled instances with supervisory features obtained from labeled instances that encompass hard-to-learn class information to enhance classification performance. Extensive experiments on semi-supervised time-series classification demonstrate the superiority of the proposed method by achieving state-of-the-art performance.