Graph-Regularized Deep Learning for EEG-Based Emotion Recognition with Psychologically-Grounded Label Structure

2026-07-11 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Health & Medical Research, Life Sciences & Biology · Depth: Expert, long

Summary

A novel graph-regularized deep learning framework enhances EEG-based emotion recognition by integrating psychological interdependencies between emotion classes. This framework conceptualizes emotions as nodes in a graph, with edges encoding proximity based on dimensional emotion theories. It employs three complementary regularization strategies—Graph Label Smoothing, Graph Laplacian, and Sliced Wasserstein Distance—to penalize model predictions that deviate from established emotion topology. Evaluated across AudioTransformer, Conformer, and DCGNN backbone architectures, the approach demonstrates architecture-agnostic benefits. Experiments on SEED-IV (4 classes) and SEED-V (5 classes) datasets show consistent improvements, achieving up to +5.42% accuracy and a 39% reduction in psychologically implausible misclassifications, thereby raising the upper bound of performance.

Key takeaway

For Machine Learning Engineers developing EEG-based emotion recognition systems, incorporating psychologically-grounded graph regularization is crucial. You should consider Sliced Wasserstein Distance with adaptive weighting as a default for its superior performance and automatic tuning, or Graph Label Smoothing for computational efficiency. This approach will not only boost your model's accuracy by up to 5.42% but also significantly reduce psychologically implausible misclassifications, enhancing clinical relevance and interpretability.

Key insights

Integrating psychological emotion topology via graph regularization significantly improves EEG-based emotion recognition accuracy and plausibility.

Principles

• Emotion classes have psychological interdependencies.
• Penalize misclassifications based on emotion proximity.
• Graph regularization is architecture-agnostic.

Method

Construct emotion graphs based on valence-arousal, then apply Graph Label Smoothing, Graph Laplacian, or Sliced Wasserstein Distance regularization, optionally with adaptive weighting.

In practice

• Use Sliced Wasserstein with adaptive weighting for best results.
• Opt for Graph Label Smoothing when computational resources are limited.
• Employ proximity violation metric to assess psychological plausibility.

Topics

• EEG Emotion Recognition
• Graph-Regularized Deep Learning
• Sliced Wasserstein Distance
• Affective Computing
• Deep Learning Architectures
• Psychological Emotion Models

Best for: AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• A Shared Valence Axis Across Modern LLMs and Human EEG: The Saturation Regularity — A shared emotional valence axis emerges independently in large language models and human brain activity, but direct alignment is counterproductive.
• Emotional regulation improves deep learning-based image classification — Emotional Regulation, using artificial subjective experience, significantly enhances deep learning image classification performance.
• EEGDancer: Dynamic Emotion Latent Space Masked Modeling with Reinforcement Learning for EEG Continuous Emotion Prediction — EEGDancer uses latent space learning, masked modeling, and reinforcement learning for continuous EEG emotion prediction.
• Adaptive Multi-Expert Graph Transformer for Interpretable EEG-Based Diagnostics — Dynamic graph transformers with adaptive expert fusion enhance interpretable, subtype-aware EEG abnormality detection.
• Fast Rates for Semi-Supervised Learning via Data-Augmentation Graph Regularization — Data augmentation induces a similarity graph, explaining fast semi-supervised learning rates tied to augmentation quality.
• Large-Language-Models-as-a-Judge in Theory-Agnostic Adaptive Metric-Alignment for Prototypical Networks in Personality Recognition — A theory-agnostic framework uses LLMs and graph networks to discover unified latent personality traits from text.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.