DAST: A Domain-Adaptive Learning Combining Spatio-Temporal Dynamic Attention for Electroencephalography Emotion...

Area of Science:

• Affective computing
• Neuroscience
• Machine learning

Background:

• Multimodal emotion recognition using electroencephalography (EEG) is prevalent.
• Existing research primarily focuses on perceived emotions, neglecting induced emotions, hindering a comprehensive understanding of two-way emotional responses.

Purpose of the Study:

• To propose a novel multimodal domain adaptive method (DAST) for EEG-based emotion recognition that integrates music as an emotional stimulus.
• To bridge the gap between induced and perceived emotions in affective computing.

Main Methods:

• Developed the DAST method utilizing spatio-temporal adaptive attention (STA-attention) for global EEG modeling.
• Employed an adaptive space encoder (ASE) for dynamic high-dimensional mapping of embeddings.
• Implemented adversarial training with a domain discriminator and ASE to learn invariant emotion representations.

Main Results:

• Extensive experiments were conducted on the DEAP dataset.
• The DAST method demonstrated an enhanced ability to explore the relationship between induced emotions (from music) and perceived emotions (from EEG).

Conclusions:

• The proposed DAST method offers a reliable approach for investigating the correlations between EEG signals and music-induced emotions.
• This research contributes to a deeper understanding of two-way emotion dynamics in affective computing.