Temporal-Frequency Co-training for Time Series Semi-supervised Learning

Semi-supervised learning (SSL) has been actively studied due to its ability alleviate the reliance of deep models on labeled data. Although existing SSL methods based pseudo-labeling strategies have made great progress, they rarely consider time-series data's intrinsic properties (e.g., temporal dependence). Learning representations by mining inherent time series recently gained much attention. Nonetheless, how utilize feature design paradigms for not explored. To this end, we propose a Time Series framework via Temporal-Frequency Co-training (TS-TFC), leveraging complementary information from two distinct views unlabeled data learning. In particular, TS-TFC employs time-domain and frequency-domain train neural networks simultaneously, each view's pseudo-labels generated label propagation in representation space are adopted guide training other classifier. enhance discriminative between categories, temporal-frequency supervised contrastive module, which integrates difficulty categories improve quality pseudo-labels. Through co-training obtained representations, is exploited enable model better learn distribution categories. Extensive experiments 106 UCR datasets show that outperforms state-of-the-art methods, demonstrating effectiveness robustness our proposed model.