<p>Time series data analysis plays an important role in numerous application domains, including medical diagnosis, solar energy forecasting, and autonomous vehicle systems. A key characteristic of such data is the scarcity of labeled samples compared to the abundance of available unlabeled data, which has driven increasing attention toward semi-supervised learning approaches for time series analysis from both research and industrial communities. The widely-used virtual adversarial training (VAT) encourages model predictions that are invariant to small input perturbations for a smooth distribution with better generalization. Although VAT performs well in vision and language tasks, directly applying it to time series classification may corrupt key trend information, reducing its effectiveness for semi-supervised learning with unlabeled data. To address the above challenges, we propose trend virtual adversarial training (tVAT), which combines trend information extracted by Gaussian blurring and Lasso-inspired adversarial perturbations to effectively leverage unlabeled data for better generalization. We further theoretically demonstrate that the perturbed input can flexibly explore sample space without introducing spike-like anomalous patterns. Empirical evaluations show tVAT’s consistent superior performance over competing baseline approaches in semi-supervised time series classification, achieving performance gains of up to 10.73%.</p>