17 June 2023 in SN Computer Science
11 February 2026 in Neurocomputing
29 May 2026 in Computers and Electrical Engineering
16 March 2026 in IEEE Signal Processing Letters
Open Set Domain Adaptation (OSDA) in non-stationary time series remains a challenging problem, particularly when the target domain contains previously unseen classes and exhibits temporal drift due to changing behaviors, sensor noise, or environmental variability. Existing adaptation methods often fail in such settings, misaligning unknown target samples and resulting in negative transfer. We propose a Confidence-Aware Latent Optimal Transport framework, named as CALOT, that performs selective domain alignment in a learned latent space. CALOT filters out low-confidence target samples—those likely to belong to unknown classes—using entropy-based confidence scoring, and applies optimal transport only to the most reliable instances. A self-regularization loss further enhances alignment stability under temporal variability. Our work achieves 90% accuracy on known classes and 80% accuracy in rejecting unknown classes, resulting in a strong H-score of 85% on the challenging non-stationary HHAR dataset. Additionally, it obtains H-scores of 76% on WISDM, 73% on HAR, 74% on EDF, and 75% on the BOILER dataset.