SCLCM: Simple Contrastive Learning based Convolutional Model for EEG Emotion Recognition

Emotions may be recognized through various sources, and while the outward expressions of emotion may be hidden, the physiological signs, such as electroencephalogram (EEG), will remain stable. One promising approach that could handle complex dependencies in EEG signals and low percentage of labeled data is self-supervised learning (SSL). However, existing approaches have complex architectures. This paper proposes a simple to understand and implement approach based on a convolutional encoder with SSL using 2 data preprocessing techniques. We also introduced a contrastive loss and batching function. The proposed approach has been evaluated on SEED and DEAP, widely used datasets for emotion classification. We achieved 87,9% and 85,8% on SEED and 93,3% and 63,1% on DEAP for subject dependent and independent splits respectively. The vector representations of the signal obtained by encoder after training on pre-text task without true labels was compressed into 2D space, and shown groups distributed on positive pairs. Another important outcome is that the encoder showed high performance during fine-tuning on a small percentage of the labeled data. In conclusion, the proposed approach reached both simplicity in architecture, effective vector representations of EEG signal and emotion classification accuracy compatible with SOTA.