Time and frequency synergy for source-free time-series domain adaptations

Notwithstanding that source-free domain adaptation (SFDA) has gained its prominence due to privacy protections of source samples in domain adaptations, very few works have been devoted to address time-series problems possessing temporal characteristics. Besides, existing works of source-free time-series domain adaptation (SFTSDA) have not exploited the potential of frequency features offering complementary information to boost the performances. This paper proposes time frequency domain adaptation (TFDA) to overcome the SFTSDA problems. TFDA fully utilizes time and frequency information via a dual branch network structure comprising time and frequency encoders. Consistencies of time and frequency domains are forced via the contrastive learning strategies in the time, frequency and time-frequency domains while applying the self-distillation concept to maintain the same. To further improve the performance, TFDA implements the uncertainty reduction strategy to combat the issue of domain shift and the curriculum learning strategy to deal with the noisy pseudo labels. Rigorous experiments with 3 time series datasets of different application domains confirm the advantage of TFDA over prior arts with noticeable margins. In addition, the theoretical analysis is provided to show the generalization bound of our approach. • This paper proposes Time Frequency Domain Adaptation (TFDA) to address the source free time-series domain adaptation problem. • TFDA puts forward the concept of a dual-branch network structure mining both temporal and frequency features. • TFDA proposes the idea of time and frequency consistencies for the source free time-series domain adaptation problem using the contrastive module and the self-distillation module. • Rigorous experiments and theoretical studies have been performed where TFDA demonstrate the most encouraging performance over the prior arts.