SDA-DDA: A Novel Transfer Learning Framework for Individualized Emotion Recognition based on EEG Signals

Objective: As one branch of human-computerinteraction, affective Brain-Computer Interfaces (aBCI) interpretand utilize electroencephalogram (EEG) signalsto achieve real-time monitoring and recognition of individualemotional states, opening new possibilities foremotion-aware technologies and applications. However,the challenge of individual differences in EEG emotiondata severely constrains the effectiveness and generalizationcapability of existing models. Method: To addressthis crucial issue, we propose a novel transfer learningframework known as Semi-Supervised Domain Adaptationwith Dynamic Distribution Alignment (SDA-DDA). Specifically,we align the marginal and conditional probabilitydistributions of the source and target domains using MaximumMean Discrepancy (MMD) and Conditional MaximumMean Discrepancy (CMMD), respectively. Subsequently, adynamic distribution adaptation algorithm is designed todynamically adjust the differences between these two distributionsduring training. In the semi-supervised domainadaptation module, we introduce a pseudo label confidencefiltering mechanism to optimize the quality of pseudo-labelgeneration and enhance the accuracy of conditional distributiondifference estimation. Result: Extensive experimentsconducted on two benchmark databases (SEED andSEED-IV) validate the reliability and stability of the model.Conclusion: Compared to existing literature, our approachachieves satisfactory results in emotion recognition underdifferent evaluation protocols, including cross-subjectand cross-session. Significance: The algorithm proposedin this study enhances the universality and reliability ofemotion recognition, promoting the development of aBCI technology and personalized applications.