dynAmiC: Dynamic Domain Adaptation with Efficient Coreset Selection

Building subject-independent models for electrocardiogram (ECG)-based emotion recognition is challenging due to substantial inter-subject variability and the high cost of utilizing large volumes of unlabeled data. While prior domain adaptation (DA) methods have mitigated distribution shifts, they typically rely on static alignment strategies and overlook data efficiency. In this work, we propose dynAmiC domain adaptation, a semi-supervised framework that dynamically balances Maximum Mean Discrepancy (MMD) and Local Structure Discriminative (LSD) losses to achieve effective global–local alignment. Furthermore, we introduce a coreset selection strategy that leverages only 1\% of the unlabeled target data while delivering performance comparable to using the entire dataset. Extensive experiments on the DREAMER and WESAD benchmarks, evaluated under leave-one-subject-out cross-validation, demonstrate that our approach consistently outperforms state-of-the-art baselines. These findings highlight dynAmiC domain adaptation as a robust and data-efficient pathway toward practical, calibration-free affective computing systems.