Local-Global Feature Fusion for Subject-Independent EEG Emotion Recognition

TL;DR

This work tackles subject-independent EEG emotion recognition by jointly modeling local channel-level dynamics and global trial-level coordination. It introduces a local–global connectivity fusion framework where per-channel differential entropy and graph-theoretic features are combined with global time-domain, spectral, and multifractal descriptors, then fused in a dual-branch transformer (MAET) with attention-based fusion and domain-adversarial regularization under a LOSO protocol on SEED-VII. The approach yields consistent cross-subject gains over local-only and classical baselines, reaching approximately $40\%$ accuracy in 7-class emotion recognition, and highlights the complementary value of spectral and multifractal global features. The results demonstrate the importance of integrating fine-grained electrode dynamics with subject-invariant coordination patterns, offering a path toward more robust and generalizable affective decoding in EEG-based applications.

Abstract

Subject-independent EEG emotion recognition is challenged by pronounced inter-subject variability and the difficulty of learning robust representations from short, noisy recordings. To address this, we propose a fusion framework that integrates (i) local, channel-wise descriptors and (ii) global, trial-level descriptors, improving cross-subject generalization on the SEED-VII dataset. Local representations are formed per channel by concatenating differential entropy with graph-theoretic features, while global representations summarize time-domain, spectral, and complexity characteristics at the trial level. These representations are fused in a dual-branch transformer with attention-based fusion and domain-adversarial regularization, with samples filtered by an intensity threshold. Experiments under a leave-one-subject-out protocol demonstrate that the proposed method consistently outperforms single-view and classical baselines, achieving approximately 40% mean accuracy in 7-class subject-independent emotion recognition. The code has been released at https://github.com/Danielz-z/LGF-EEG-Emotion.

Figures (5)

• Figure 1: Overall workflow of the proposed pipeline on the SEED-VII dataset. Preprocessed EEG signals were used to extract local channel-wise and global trial-level features, which were subsequently integrated via an MAET-based fusion model and evaluated under a LOSO protocol.
• Figure 2: Architecture of the MAET backbone: multi-view embeddings for the EEG feature vector (local dynamics and node-level connectivity) and the auxiliary feature vector, followed by transformer blocks and a single emotion classification head.
• Figure 3: Mean normalized confusion matrix averaged over LOSO folds for 7-class emotion recognition.
• Figure 4: Per-class precision, recall, and F1-score (mean over LOSO folds).
• Figure 5: Subject-wise accuracy under LOSO evaluation. The dashed line denotes the mean accuracy across subjects.