ADAPTER: Auto-Augmentation for Emotion Recognition in EEG - A Class and Subject Invariant Approach

Data Augmentation is essential to machine learning pipelines, particularly for developing robust models in emotion recognition tasks using EEG signals. The low signal-to-noise ratio of EEG signals and limitations on collecting large EEG datasets make data augmentation one of the easiest and most effective techniques for developing real-world emotion recognition models. It addresses class imbalance, distribution shifts, and lack of data, allowing the model to generalise and reduce overfitting. This paper introduces ADAPTER (Automatic Data Augmentation and PerTubation for Emotion Recognition), a novel automatic data augmentation strategy for emotion recognition using EEG signal classification. ADAPTER finds the best policy of geometric transformations across multiple classes in the classification dataset, such as SEED, without manually selecting from the pool of different transformations. It is tailored to be class and subject-invariant, ensuring that the resulting models are robust and generalisable across various tasks and subjects. The proposed technique outperforms the existing state-of-the-art model for SEED by 5% in terms of accuracy and 2% macro F1-score. ADAPTER showcases the potential of a flexible and adaptable data augmentation approach in enhancing the performance of EEG-based emotion recognition models while maintaining interoperability and control over the augmentation process.