PR²DM: Position-aware robust reconstruction with diffusion model for emotion recognition From EEG

Emotion recognition is significant in domains like smart education, mental health assistance, and brain-computer interface. Electroencephalography-based emotion recognition methods provide an objective approach to evaluating emotional states by analyzing the brain's electrophysiological activity. In this research, we aim to address the challenges of EEG data scarcity and noise that vary depending on the electrode position, both of which significantly impact the quality of emotion recognition.

We propose position-aware robust reconstruction with diffusion model (PR²DM) for emotion recognition from EEG. PR²DM integrates electrode positional encoding with features in each level and fuses various temporal scale features, reconstructing noise-independent features from the corrupted features to enrich the training sample. Additionally, we propose a spatial-temporal adaptive convolutional network that adaptively adjusts parameters based on EEG features for more precise classification.

The developed framework is validated on DEAP and DREAMER datasets, demonstrating excellent performance in both valence and arousal classification tasks, achieving 95.01% and 95.62% accuracy on DEAP, and 84.77% and 87.30% accuracy on DREAMER, respectively. A comparative analysis of EEG topographic maps further confirms the model's effectiveness in reconstructing noise-independent EEG features. Ablation experiments further explored the contribution of each component to model performance, proving that our proposed method effectively enhances the performance of emotion recognition.

The proposed PR²DM mitigates the impact of data scarcity in EEG signals and inconsistent noise distribution across electrode positions, suggesting improvements in emotion recognition accuracy based on EEG signals.