Trend and Order Features for Semi-Supervised Time-Series Classification via Multitask Learning.

Area of Science:

• Machine Learning
• Data Science
• Time Series Analysis

Background:

• Multitask learning with pretext tasks improves time-series classification, especially with scarce labeled data.
• Effective feature extraction from raw time series is crucial for multitask learning success.

Purpose of the Study:

• To propose a novel semi-supervised time-series classification method using multitask learning, named TOFL.
• To introduce trend and order features for enhanced classification performance.

Main Methods:

• Developed a self-sequence order prediction (SOP) pretext task to learn temporal order relations.
• Designed a gradual trend fusion (GTF) block to extract high-quality trend features for the SOP task.
• Theoretically analyzed uniform stability and generalization error of the proposed TOFL framework.

Main Results:

• TOFL demonstrated high competitiveness against state-of-the-art (SOTA) supervised and semi-supervised methods.
• The proposed method closely matched or surpassed SOTA accuracy on 128 UCR datasets and three real-world datasets.
• Source code and data are publicly available for reproducibility.

Conclusions:

• TOFL offers a robust and effective approach for semi-supervised time-series classification.
• The combination of SOP and GTF enables superior feature representation for time-series data.
• The method shows significant potential for practical applications requiring accurate time-series classification with limited labels.