Time Series Classification Based on Supervised Contrastive Learning and Homoscedastic Uncertainty

• Published in: IEEE transaction on neural networks and learning systems Vol. PP; pp. 1 - 14
• Main Authors: Zhang, Tao, Li, Ke, Wang, Shaofan
• Format: Journal Article
• Language: English
• Published: United States IEEE 24.09.2025
• Subjects: Data augmentation; Feature extraction; Frequency-domain analysis; Learning systems; Multitask learning (MTL); multivariate time series classification (MTSC); Noise measurement; Representation learning; supervised contrastive learning (SCL); Time series analysis; Time-domain analysis; Training; Uncertainty
• ISSN: 2162-237X; 2162-2388; 2162-2388
• DOI: 10.1109/TNNLS.2025.3607901

In recent years, contrastive learning (CL) frameworks have been widely applied to multivariate time series classification (MTSC) tasks. However, existing methods lack task-specific guidance, leading to limitations in fully capturing the complex dynamics and invariant representations in time series data. Motivated by the auxiliary tasks in multitask learning (MTL) and to fully utilize the rich frequency-domain information of time series data, we propose a novel time series classification framework, uncertainty-based time-frequency supervised CL (U-TFSCL). This framework uses SCL in the time and frequency domains and time-frequency consistency as auxiliary tasks to improve the primary task of using only instance-level labels for time series classification. Furthermore, inspired by the homogeneous uncertainty in MTL, we derive a novel uncertainty loss function, which automatically adjusts the weights according to the degree of uncertainty of different tasks to optimize the learning and prediction process of the model. The proposed framework is evaluated on MTSC tasks, including human activity recognition (HAR), air writing, and gesture recognition. In addition, we create a human-drone interaction (HDI) dataset consisting of 20 subjects and conduct real-world experiments to evaluate the proposed framework. The extensive experiments conducted in various settings verify the effectiveness of the proposed framework.