Data augmentation for enhancing EEG-based emotion recognition with deep generative models

Area of Science:

• Neuroscience
• Machine Learning
• Data Science

Background:

• Emotion recognition from electroencephalography (EEG) faces challenges due to data scarcity, hindering the development of accurate machine learning and deep neural network models.
• Deep generative models offer potential solutions for augmenting limited EEG datasets.

Purpose of the Study:

• To address the data scarcity problem in EEG-based emotion recognition by proposing novel data augmentation methods.
• To enhance the performance of affective models through improved EEG training datasets.

Main Methods:

• Proposed three data augmentation methods: conditional Wasserstein GAN (cWGAN), selective VAE (sVAE), and selective WGAN (sWGAN).
• Employed full and partial usage strategies for data augmentation, with partial usage selecting only high-quality generated data.
• Generated realistic EEG data in power spectral density and differential entropy forms for augmentation.

Main Results:

• Systematic experiments on SEED and DEAP EEG datasets demonstrated the effectiveness of the proposed methods.
• Generative model-based augmentation outperformed existing methods like conditional VAE, Gaussian noise, and rotational augmentation.
• Optimal performance was achieved when the number of generated data points was less than 10 times the original dataset size.

Conclusions:

• The selective WGAN (sWGAN) method significantly enhances the performance of EEG-based emotion recognition models.
• Data augmentation using generative models is a viable strategy to overcome data scarcity in affective computing.