A Hierarchical Meta-Learning Driven Adaptive Brain Region Flow Fusion Method for EEG-Based Emotion Recognition

EEG (electroencephalography)-based emotion recognition has garnered substantial attention in affective computing. However, existing state-of-the-art methods still face two major challenges: 1) It is difficult to strike a balance between a priori physiological consistency and data-driven flexibility in the functional parcellation of brain regions; 2) Ineffective synergistic adaptation to individual differences and local-global feature correlations. To tackle these, we propose HMABRF, a hierarchical meta-learning-driven adaptive brain region flow fusion method for EEG-based emotion recognition. Guided by affective neuroscience priors, we design an adaptive brain region flow partitioning mechanism. We begin by using self-supervised contrastive learning to enhance the semantic consistency of channel features, and then employ a differentiable Gumbel-Softmax function to perform dynamic clustering of sub-flows; in addition, based on a hierarchical meta-learning framework, we develop LTIM (Local Tuned Interaction Module) and GTCM (Global Temporal Calibration Module). LTIM enhances local multi-scale feature interaction through its built-in Bilateral Feature Interaction and Enhancement (BIE), specifically adapting to individual differences in local details; GTCM leverages Mamba and the Contextual Temporal Interaction Relay (CIR) to capture long-range temporal dependencies and accurately adapt to subject-specific global long-term feature patterns across the entire sequence, while using a self-supervised alignment loss to enhance the semantic consistency between local and global features. The experimental results show that HMABRF outperforms existing state-of-the-art methods on the SEED and SEED-IV datasets.