Contrastive Twin Latent Informer Based on Autoencoder for EEG Emotion Recognition

Abstract

Electroencephalography (EEG) is widely used in emotion recognition due to its reliability and noninvasiveness. However, challenges as signal noise and individual variability in EEG data present significant limitations for real-world applications. Inspired by recent neuroscience advances, we propose a contrastive twin latent informer (CTLIN) model for EEG-based emotion recognition and signal generation. The architecture employs a symmetric autoencoder (AE) that synergistically integrates PropSparse self-attention mechanisms with EEG feature extraction, incorporating a novel twin latent permutation (LP) design. By disentangling EEG features into subject and emotion latent representations through contrastive learning, the model demonstrates enhanced capability in capturing generalizable biomarkers. Experimental results show CTLIN’s state-of-the-art performance in zero-shot classification tasks. It significantly surpasses existing models in both subject identification and emotion classification across three benchmark datasets [SEED, SEED-IV, and a self-collected (SC) dataset], achieving subject classification accuracies of 98.63%, 99.13%, and 98.70%, respectively, with corresponding emotion classification accuracies of 77.92%, 53.79%, and 64.41%. Furthermore, a novel paradigm for EEG signal generation and data augmentation is established through the proposed LP approach, providing practical solutions to enhance emotion recognition.