MB-MSTFNet: A Multi-Band Spatio-Temporal Attention Network for EEG Sensor-Based Emotion Recognition.

Area of Science:

• Affective computing
• Neuroscience
• Machine learning

Background:

• Emotion recognition from electroencephalogram (EEG) is crucial for human-machine interaction.
• Existing methods face challenges in spatio-temporal feature fusion and integrating multi-band and brain-region information.
• Accurate emotion recognition requires sophisticated analysis of complex EEG sensor signals.

Purpose of the Study:

• To propose MB-MSTFNet, a novel framework for enhanced EEG-based emotion recognition.
• To address the limitations in spatio-temporal feature fusion and cross-band/brain-region integration in EEG analysis.
• To develop a sensor-driven model for efficient and accurate real-time emotion recognition.

Main Methods:

• Constructed a 3D tensor to encode band-space-time correlations in EEG sensor data.
• Employed a multi-scale CNN-Inception module for hierarchical spatial feature extraction.
• Utilized Bi-directional GRUs (BiGRUs) and multi-head self-attention for temporal dependency and channel weighting.

Main Results:

• Achieved 96.80% valence and 98.02% arousal accuracy in binary classification on the DEAP dataset.
• Attained 92.85% accuracy for four-class emotion classification.
• Ablation studies confirmed the significant performance enhancement from feature fusion, bidirectional temporal modeling, and multi-scale mechanisms.

Conclusions:

• MB-MSTFNet effectively integrates spatio-temporal dynamics and multi-band interactions of EEG sensor signals.
• The proposed framework significantly advances affective computing for real-time emotion recognition.
• The model demonstrates superior performance in capturing complex patterns within EEG data for emotion analysis.