Group Resonance Network: Learnable Prototypes and Multi-Subject Resonance for EEG Emotion Recognition

Electroencephalography(EEG)-basedemotionrecognitionre- mains challenging in cross-subject settings due to severe inter-subject variability. Existing methods mainly learn subject-invariant features, but often under-exploit stimulus-locked group regularities shared across sub- jects. To address this issue, we propose the Group Resonance Network (GRN), which integrates individual EEG dynamics with offline group resonance modeling. GRN contains three components: an individual en- coder for band-wise EEG features, a set of learnable group prototypes for prototype-induced resonance, and a multi-subject resonance branch that encodes PLV/coherence-based synchrony with a small reference set. A resonance-aware fusion module combines individual and group-level rep- resentations for final classification. Experiments on SEED and DEAP under both subject-dependent and leave-one-subject-out protocols show that GRN consistently outperforms competitive baselines, while abla- tion studies confirm the complementary benefits of prototype learning and multi-subject resonance modeling.

💡 Research Summary

The paper tackles one of the most persistent challenges in EEG‑based affective computing: the severe inter‑subject variability that degrades performance when models are transferred to unseen participants. While many recent works focus on domain adaptation, adversarial alignment, or learning subject‑invariant embeddings, they largely treat each subject as an isolated domain and ignore the stimulus‑locked regularities that emerge across subjects exposed to the same emotional video or audio clip. Neuroscience literature, however, reports measurable inter‑subject synchrony—quantified by Phase Locking Value (PLV) and spectral coherence—when multiple individuals experience the same affective stimulus. Leveraging this insight, the authors propose the Group Resonance Network (GRN), a novel architecture that jointly models (i) individual EEG dynamics, (ii) learnable group prototypes that capture the latent structure of the whole cohort, and (iii) explicit multi‑subject resonance computed against a small reference set.

Architecture Overview

1. Individual Encoder – Raw EEG segments (C channels × T time points) are first transformed into band‑wise features across B conventional frequency bands (δ, θ, α, β, γ). These channel‑band matrices are fed into a lightweight CNN or Transformer backbone, producing a d‑dimensional embedding F that encodes subject‑specific temporal‑spatial patterns.
2. Learnable Group Prototypes – Instead of fixing a single “average subject”, the model maintains M trainable prototype vectors {p₁,…,p_M} ∈ ℝ^d. For a given F, cosine similarities to all prototypes are turned into soft attention weights αₘ, and a prototype‑induced resonance vector R = Σₘ αₘ pₘ is formed. R reflects which group‑level pattern the current sample aligns with most closely.
3. Multi‑Subject Resonance Tensor – A small reference pool S = {X^(k)}_{k=1}^{K_r} is sampled from the training subjects (never from the held‑out subject in LOSO). For each reference, the PLV matrix (phase synchrony) and the coherence matrix (frequency‑domain coupling) between the current sample and the reference are computed, yielding a stacked tensor M ∈ ℝ^{K_r × C × C × 2}. A shallow 2‑D CNN with pooling (ResEnc) compresses M into a d‑dimensional vector G, which encodes the structural synchrony between the target subject and the reference cohort under the same stimulus.

Resonance‑Aware Fusion
The three embeddings (F, R, G) are combined by explicitly modeling both difference and commonality terms: (F−R), (F−G) capture subject‑specific deviations from group patterns, while element‑wise products F⊙R and F⊙G capture shared components. All four terms are concatenated and fed to a multilayer perceptron (MLP) that outputs the final representation for a softmax classifier. This design forces the network to preserve individual cues while exploiting stimulus‑locked group information.

Training Objective
The loss consists of the standard cross‑entropy classification term L_cls plus a prototype regularizer L_proto weighted by λ. The regularizer encourages the individual embedding to stay close to high‑attention prototypes, stabilizing prototype learning.

Experimental Setup
Two widely used affective EEG benchmarks are used: SEED (three emotion classes: negative, neutral, positive) and DEAP (binary valence and arousal). Both subject‑dependent (SD) and subject‑independent (leave‑one‑subject‑out, LOSO) protocols are evaluated. Hyper‑parameters: embedding dimension d=256, number of prototypes M=8, reference pool size K_r=3, Adam optimizer (lr=1e‑4, weight decay=1e‑4), batch size 64, up to 80 epochs with early stopping. PLV is derived from analytic phase via Hilbert transform; coherence is computed from Welch PSD and averaged across bands.

Results
GRN consistently outperforms all baselines (DGCNN, ST‑DADGAT, FCANet, LATN, DVIE‑Net) on both datasets and protocols. For SEED LOSO, GRN reaches 87.90 % accuracy versus 86.34 % for the previous best (DVIE‑Net). On DEAP, GRN achieves 90.35 % (valence) and 89.40 % (arousal), again surpassing the prior state‑of‑the‑art.

Ablation Studies

• Individual only (no R, no G): 84.60 %
• Prototypes only (R, no G): 86.40 %
• Multi‑subject resonance only (G, no R): 86.85 %
• Full GRN (R + G): 87.90 %
• Full GRN without prototype regularizer: 87.35 %

These results demonstrate that both prototype learning and multi‑subject resonance contribute complementary gains. Sensitivity analysis shows performance is stable across K_r ∈ {1,3,5} and M ∈ {4,8,12}.

Interpretability
Confusion matrices reveal that GRN maintains balanced predictions across the three SEED emotions under LOSO, indicating reduced inter‑subject bias, while under SD it sharpens diagonal dominance, reflecting stronger capture of subject‑specific nuances. Training/validation curves show smooth convergence without over‑fitting.

Conclusion and Impact
GRN introduces a dual‑level group modeling paradigm that unifies learnable latent prototypes with concrete neurophysiological synchrony measures. By fusing these with individual EEG embeddings through a resonance‑aware module, the network achieves superior cross‑subject generalization while preserving subject‑specific details. The approach is compatible with modern transformer‑style EEG backbones and can be readily extended to other BCI tasks, real‑time affective computing, or multimodal affect detection. Code is publicly released, facilitating reproducibility and future research.