FAT 🧠 Fourier Adjacency Transformer for EEG Emotion Recognition

(Early accept by MICCAI 2025) This is the official implementation of the paper:

A Novel Fourier Adjacency Transformer for Advanced EEG Emotion Recognition

📄 Paper: Click here to read the paper

Abstract EEG emotion recognition faces significant hurdles due to noise interference, signal nonstationarity, and the inherent complexity of brain activity which make accurately emotion classification. In this study, we present the Fourier Adjacency Transformer, a novel framework that seamlessly integrates Fourier-based periodic analysis with graph-driven structural modeling. Our method first leverages novel Fourier-inspired modules to extract periodic features from embedded EEG signals, effec tively decoupling them from aperiodic components. Subsequently, we em ploy an adjacency attention scheme to reinforce universal inter-channel correlation patterns, coupling these patterns with their sample-based counterparts. Empirical evaluations on SEED and DEAP datasets demon strate that our method surpasses existing state-of-the-art techniques, achieving an improvement of approximately 6.5% in recognition accu racy. By unifying periodicity and structural insights, this framework of fers a promising direction for future research in EEG emotion analysis.

Figure: Overview of the FAT model architecture.

Requirements:

• Python 3.10
• PyTorch 2.2.2
• CUDA 11.8+ (if using GPU)

Datasets:

Please refer to the methodological division in the paper

• SEED

  A multi-subject emotion recognition dataset with 3-class labels (positive, neutral, negative) recorded via 62-channel EEG.

• SEED-IV

  A 4-class version of SEED, using similar experimental settings but with an expanded emotion label set.

• SEED-V

  A 5-class EEG emotion dataset with higher emotional resolution. Includes: joy, sadness, anger, fear, neutral.

• SEED-VII

  A 7-class dataset for fine-grained emotion recognition using 62-channel EEG across multiple sessions.

• DEAP

  Dataset for emotion analysis using EEG and peripheral physiological signals (32 channels), includes video-based stimuli.

Citation

If you use this repository or model in your research, please cite 😊: