Time-frequency based multi-task learning for semi-supervised time series classification

Semi-supervised learning is crucial for alleviating labeling burdens in time series classification. Most of the existing semi-supervised time series classification methods extract patterns from the time domain, ignoring the time-frequency domain and the latent feature space shared by the labeled and unlabeled samples. For that, a Multi-task learning scheme based on Time-Frequency mining for semi-supervised time series Classification (MTFC) is proposed. First, we design a series of unsupervised tasks for capturing time-frequency information. Considering the consistency between labeled and unlabeled data, we then employ a multi-task learning framework to learn their common features. Meanwhile, we theoretically analyze the proposed semi-supervised classification method and provide a novel generalization result for the MTFC. Extensive experiments on multiple time series datasets demonstrate that our MTFC can effectively improve the performance of semisupervised classification and achieve state-of-the-art results. (c) 2022 Elsevier Inc. All rights reserved.

Automated summary

A multi-task learning scheme based on Time-Frequency mining is proposed for semi-supervised time series classification. It captures time-frequency information and learns common features through a multi-task learning framework, improving classification performance and achieving state-of-the-art results.