Emotion Recognition from DEAP: Feature Engineering & Baselines

This repository supports a preprint study on emotion recognition using the DEAP EEG dataset, focusing on carefully engineered signal-level features and clean baselines.

Key elements:

• Balanced label construction for valence and arousal using carefully chosen thresholds (e.g., 5.05 and 5.15) to avoid class imbalance.
• Multiple feature families with full code:
  • Wavelet Energy Bands (WEB)
  • Hilbert–Huang Transform Entropy (HHTE)
  • Hilbert Spectrum Energy (HSE)
  • Wavelet-based Differential Entropy (DE)
• Detailed spatiotemporal mapping pipeline:
  • Scaling and normalization of features
  • Temporal segmentation of long EEG windows (e.g., 60×128)
  • Projection of 32 EEG channels to an 8×8 2D grid for CNN input
• Final 3D representations formatted for 3D-CNN models, preserving both spatial and temporal information.

This project documents the feature-engineering and representation learning foundations that later evolved into the more advanced dual-path and GCP-based models.