CompleMatch: Boosting Time-Series Semi-Supervised Classification With Temporal-Frequency Complementarity.

Area of Science:

• Machine Learning
• Data Science
• Signal Processing

Background:

• Semi-Supervised Classification (SSC) leverages unlabeled data to improve model performance when labeled samples are scarce.
• Existing time series SSC methods primarily rely on temporal dependencies, which can be sensitive to noise and may miss global feature periodicity.

Purpose of the Study:

• To introduce CompleMatch, a novel time series SSC framework that utilizes complementary information from both temporal and frequency domains.
• To enhance the learning from unlabeled data by integrating diverse data representations.

Main Methods:

• CompleMatch employs a co-training paradigm with two simultaneously trained deep neural networks, one for time-domain and one for frequency-domain views.
• Pseudo-labels generated via label propagation guide the training of each network, exploiting the complementary nature of temporal-frequency representations.
• A temporal-frequency contrastive learning module integrates supervised and self-supervised signals to improve pseudo-label quality and representation discriminability.

Main Results:

• CompleMatch significantly outperforms state-of-the-art methods in time series SSC tasks.
• Ablation studies and visualizations confirm the effectiveness of the proposed temporal-frequency complementary learning strategy.
• The framework demonstrates enhanced robustness and performance, particularly under conditions of limited labeled data.

Conclusions:

• The proposed CompleMatch framework effectively leverages complementary temporal and frequency information for robust time series SSC.
• Integrating diverse data representations and contrastive learning enhances model performance and discriminative power.
• CompleMatch offers a promising advancement for semi-supervised learning in time series analysis.