A temporal-spatial feature fusion network for emotion recognition with individual differences reduction.

Area of Science:

• Neuroscience
• Machine Learning
• Affective Computing

Background:

• EEG-based emotion recognition faces challenges due to individual variability and limitations of conventional time-series models, often resulting in suboptimal cross-subject performance.
• Traditional methods extract and fuse spatial and temporal features, but struggle to generalize across different subjects.

Purpose of the Study:

• To propose a novel network, the Time-Space Emotion Network (TSEN), for enhanced EEG-based emotion recognition.
• To address the limitations of cross-subject emotion recognition by effectively fusing spatiotemporal information.

Main Methods:

• TSEN incorporates a Convolutional Block Attention Module (CBAM) for weighted spatial feature extraction.
• A residual block with Switchable Whitening (SW) is used to enhance network stability and domain adaptation.
• Temporal Convolutional Networks (TCN) are employed for efficient and accurate temporal feature extraction, maintaining a lightweight model.

Main Results:

• Experiments on the DEAP dataset showed average accuracy for arousal prediction of 0.7032 (variance 0.0876) and an F1 score of 0.6843.
• Valence prediction achieved an accuracy of 0.6792 (variance 0.0853) with an F1 score of 0.6826.
• TSEN demonstrated high accuracy and low variance in cross-subject emotion prediction.

Conclusions:

• TSEN effectively reduces individual differences in EEG, leading to improved cross-subject emotion recognition accuracy.
• The network's smaller parameter count allows for faster execution, making it computationally efficient.
• TSEN presents a promising approach for robust and efficient EEG-based emotion recognition.