graph-regularized-eeg-emotion-recognition

name	graph-regularized-eeg-emotion-recognition
description	Graph-regularized learning framework for EEG-based emotion recognition using psychological emotion topology. Conceptualizes emotions as nodes in a graph with edges encoding proximity based on dimensional emotion theories. Use when building EEG emotion classifiers, affective BCI systems, or applying graph regularization to psychological classification tasks.

Graph-Regularized EEG-Based Emotion Recognition

Description

A graph-regularized learning framework that improves EEG-based emotion recognition by modeling emotions as interconnected nodes in a graph rather than isolated labels. Edges encode psychological proximity based on dimensional emotion theories, penalizing predictions that deviate from established emotion topology.

Paper: Graph-Regularized Deep Learning for EEG-Based Emotion Recognition with Psychologically-Grounded Label Structure (arXiv:2607.07773, July 2026)

Activation Keywords

• graph-regularized emotion recognition
• EEG emotion graph
• psychological label structure
• emotion topology learning
• affective BCI graph
• 脑电情绪图正则化
• graph label smoothing emotion
• Wasserstein emotion classification

Core Insight

Traditional deep learning for EEG emotion recognition treats emotion classes as isolated labels, ignoring psychological interdependencies. This framework recognizes that emotions exist in a continuous dimensional space (e.g., valence-arousal), and misclassifications between psychologically adjacent emotions are more acceptable than distant ones.

Three Regularization Strategies

1. Graph Label Smoothing (Lowest Complexity)

• Intuitive soft labeling based on emotion graph proximity
• Instead of one-hot labels, distribute probability mass across neighboring emotion nodes
• Computationally lightweight, easy to implement

2. Graph Laplacian Commuting Distance (Medium Complexity)

• Spectral graph theory approach
• Uses graph Laplacian eigendecomposition to compute commuting distances between emotion nodes
• Penalizes predictions based on spectral distance in emotion space

3. Sliced Wasserstein Distance (Highest Complexity)

• Optimal transport on emotion graph
• Computes minimum "cost" to transform predicted distribution to true distribution on graph
• Most theoretically grounded but computationally expensive

Results

Metric	Improvement
Best accuracy gain	+5.42%
Reduction in implausible misclassifications	39%
Architecture-agnostic	Works with AudioTransformer, Conformer, DCGNN
Datasets	SEED-IV (4 classes), SEED-V (5 classes)

Implementation Guide

Step 1: Build Emotion Graph

import numpy as np
from scipy.spatial.distance import pdist, squareform

# Define emotion positions in valence-arousal space
emotion_coords = np.array([
    [0.8, 0.6],   # happy
    [-0.7, 0.3],  # sad
    [0.3, 0.8],   # excited
    [-0.5, -0.4], # afraid
])

# Compute pairwise distances as edge weights
distances = squareform(pdist(emotion_coords))
adjacency = np.exp(-distances / sigma)  # RBF kernel

Step 2: Apply Regularization

# Graph Label Smoothing
def graph_label_smooth(one_hot, adjacency, alpha=0.1):
    """Smooth labels using graph adjacency."""
    smooth = (1 - alpha) * one_hot + alpha * adjacency @ one_hot
    return smooth / smooth.sum()

# Graph Laplacian Regularization
def laplacian_regularizer(predictions, laplacian):
    """Penalize predictions that violate graph smoothness."""
    return torch.sum(predictions.T @ laplacian @ predictions)

# Sliced Wasserstein Distance
def swd_on_graph(pred_dist, true_dist, projections=100):
    """Compute sliced Wasserstein distance on emotion graph."""
    # Project distributions onto random directions
    # Compute 1D Wasserstein distance for each projection
    # Average over all projections

Step 3: Integrate with Backbone

# Works with any backbone architecture
# AudioTransformer, Conformer, DCGNN, etc.
# Add regularization term to loss:
# total_loss = cross_entropy_loss + lambda * graph_regularization

Pitfalls

• Emotion graph construction: The quality of results depends heavily on how emotions are positioned in the dimensional space. Use validated psychological models (e.g., circumplex model of affect)
• Regularization strength: Lambda must be tuned - too strong oversmooths, too weak provides no benefit
• Dataset size: Graph regularization shows most benefit on smaller datasets where overfitting is a concern
• Architecture-agnostic: Framework works with any backbone, but benefits may vary

When to Use

• EEG-based emotion recognition with limited labeled data
• Affective brain-computer interface development
• Any classification task where labels have known topological/psychological relationships
• Reducing clinically implausible misclassifications in mental health monitoring

Resources

• Paper: https://arxiv.org/abs/2607.07773
• PDF: https://arxiv.org/pdf/2607.07773v1
• Code: Will be released (paper states "Code will be released")
• Datasets: SEED-IV, SEED-V