Multi-Source Reinforced Selective Domain Adaptation for Cross-Subject and Cross-Session EEG-Based Emotion Recognition

In recent years, transfer learning methods have found widespread applications in emotion recognition, especially in domains with limited labeled data. The significant distribution disparity between source and target domains in transfer learning often leads to detrimental negative transfer phenomena. To mitigate this issue, the judicious selection of source domain data has proven to be an effective strategy. However, traditional training data selection methods typically require predefined thresholds, which are not easy to set, and the model is trained separately from the training data selection process. In an effort to address this gap, we introduce the Multi-Source Reinforced Selective Domain Adaptation (MSRSDA) model. The MSRSDA model comprises three components: a data augmentation module, a data selector based on the actor-critic framework, and a domain adaptation module. The data augmentation module generates new samples for the data selector to consider. The data selector, acting as an actor, performs actions to select high-quality source domain data for optimization of the domain adaptation model. Additionally, we propose an innovative domain adaptation model that provides reward to update the data selector. To the best of our knowledge, our work represents an initial attempt to apply reinforcement learning to EEG-based data selection. Extensive experiments show that the MSRSDA model achieves the best current results for emotion recognition tasks on multiple benchmark datasets.