Abstract
Electroencephalography (EEG)-based emotion recognition is a critical component of advanced human-computer interaction, yet it is frequently hampered by data scarcity and significant class imbalances. Traditional Generative Adversarial Networks (GANs) often struggle with minority classes, tending to mimic limited observed distributions rather than exploring the broader emotional subspace. This paper introduces the Emotional Subspace Constrained GAN (ESC-GAN), a novel framework that shifts from simple data mimicry to a targeted "editing" paradigm. By transforming well-represented EEG signals into under-represented emotional subspaces and employing diversity- and boundary-aware constraints, ESC-GAN significantly improves classification accuracy and adversarial robustness.
Introduction
The effectiveness of deep learning in EEG emotion recognition is often bottlenecked by the labor-intensive nature of data collection, leading to small and imbalanced datasets. In many public benchmarks, certain emotional states (e.g., negative valence) are under-represented. When standard GANs are used to augment these minority classes, they often suffer from mode collapse—generating redundant samples that fail to capture the true variance of the emotional state. This leaves recognition models biased toward majority classes and fragile near decision boundaries.
The ESC-GAN Framework
ESC-GAN addresses these limitations by introducing an "EEG editing" paradigm. Instead of generating synthetic signals from random noise, the generator takes a high-quality "reference" signal from a majority class and a target emotional label. It then "edits" the reference signal to shift its emotional content while preserving the underlying physiological structure of the EEG.
This approach ensures that generated samples are grounded in realistic brain activity patterns. To ensure the quality of these edits, the framework utilizes a dual-discriminator system: one discriminator evaluates the "realness" of the signal, while another verifies that it accurately reflects the target emotion.
Key Loss Functions
The core innovation of ESC-GAN lies in its specialized loss functions designed to maximize the utility of augmented data.
Diversity-Aware Loss ($L_G^{(div)}$)
To prevent the generator from producing near-identical samples, a diversity constraint is applied. Using a Radial Basis Function (RBF) kernel, this loss function encourages the generator to maximize the distance between pairs of generated samples in the feature space, ensuring broad coverage of the target emotional subspace.
Boundary-Aware Loss ($L_G^{(bou)}$)
Classification errors predominantly occur near the decision boundaries between emotional states. ESC-GAN intentionally generates samples in these "confusing" regions. By minimizing the KL divergence between the discriminator’s prediction and a uniform distribution, the model is forced to produce samples that lie on the edge of the classifier's certainty.
Experimental Evaluation
The framework was rigorously tested on three benchmark datasets: DEAP, AMIGOS, and SEED.
- Performance: ESC-GAN achieved state-of-the-art results, with 96.33% accuracy on DEAP and 94.40% on AMIGOS.
- Imbalance Mitigation: In extreme "hard-case" scenarios where 90% of minority samples were removed, ESC-GAN boosted minority class accuracy by over 20%, demonstrating its ability to effectively populate empty emotional subspaces.
- Adversarial Robustness: By "battle-hardening" the classifier at the decision boundaries, the model showed superior resistance to adversarial attacks. Under Fast Gradient Sign Method (FGSM) attacks, the ESC-GAN-trained model maintained 71.22% accuracy, significantly outperforming the 66.15% baseline.
Conclusion
ESC-GAN represents a significant shift in EEG data augmentation, moving from passive mimicry to active subspace exploration. By focusing on diversity and decision boundaries, the framework not only solves the data scarcity problem but also produces more robust and reliable emotion recognition models. Future research will explore the application of this editing paradigm to multi-modal physiological signals and more complex emotional models.
Summary updated