EEG Coupled Scale-Invariant Dynamics for Emotion Recognition: A Domain Adaptation Approach

Abstract

In electroencephalogram (EEG)-based emotion recognition, traditional static univariate models often struggle to capture the complex, scale-free dynamics inherent in multivariate neural signals, which hampers generalization across subjects. To address this, we introduce a novel stochastic framework based on operator multifractional Levy stable motion (omLsm) and stochastic differential equations (SDE). This framework effectively captures the dynamic scale-free properties of EEG signals and assesses their local cross-scaling characteristics, revealing dynamic fractal connectivity that correlates with various emotional states. The rationale behind our approach lies in the shared scale-free properties and affective cognitive attributes observed across different subjects within the same emotion categories. Local cross-scaling characteristics expose commonalities in the spatio-temporal and spectral domains, facilitating more robust emotion recognition through a multivariate lens. Furthermore, our framework incorporates domain adaptation strategies that enhance model performance across diverse subject populations. Our results indicate significant differences in scale-free connectivity associated with emotional states, reflecting clear advantages over static univariate approaches. Notably, our detection method achieves maximum accuracy of 98.00 for dominance and 98.41 for arousal recognition, respectively, using the DREMER and DEAP datasets and cross-dataset experiments, demonstrates impressive generalization capabilities of the proposed model. This signifies our method's effectiveness for practical applications in emotion recognition.