Bridging Neural Topology and Affective Computing: Graph Attention for EEG Emotion Recognition

TL;DR

This paper reviews how graph attention networks improve emotion recognition from EEG data by modeling brain connectivity patterns.

Contribution

The paper introduces a standardized framework to improve reproducibility in graph-based EEG emotion recognition models.

Findings

Graph attention mechanisms preserve brain topology while emphasizing emotion-relevant regions.

Standardized preprocessing and validation protocols enhance model performance and reproducibility.

Future research should focus on model efficiency and multimodal integration for better generalization.

Abstract

Electroencephalography (EEG) offers high temporal resolution and strong physiological validity for emotion recognition. However, complex spatial organization and inter-subject variability present major modeling challenges. Graph-based spatial attention mechanisms have emerged as a key solution, preserving brain topological priors while adaptively emphasizing emotion-relevant regions and connections. This review summarizes advances in graph convolutional networks (GCN) and graph attention networks (GAT), covering representative studies under both subject-dependent and subject-independent settings. In architectural innovations, this paper critically evaluates the implicit impact of experimental factors, including preprocessing pipelines and validation protocols, on performance, and proposes a standardized framework to enhance reproducibility. Existing research demonstrates progressive transitions from static to dynamic graphs and from single-domain to multimodal fusion guided by physiological priors. Future research is expected to focus on enhancing model efficiency, strengthening neurophysiological alignment, integrating multimodal information and enhancing subject-independent generalization, and extending applications to affective neuroscience and clinical contexts. These developments collectively drive EEG-based emotion recognition toward more efficient, interpretable, and translationally valuable affective computing systems.

Figures