Classification-aware conditional diffusion-augmented criss-cross graph transformer for EEG-based emotion recognition

Abstract

Reliable electroencephalogram (EEG)-based emotion recognition remains challenging because labeled recordings are limited, noisy, non-stationary, and strongly subject dependent. This paper proposes a classification-aware conditional diffusion-augmented criss-cross graph transformer for raw EEG emotion recognition. Unlike generation-centered augmentation methods, the proposed framework couples EEG generation and classification through asymmetric alternating optimization. Detached multi-step synthetic samples expand the classifier training distribution, while differentiable one-step clean estimates receive feedback from a temporarily fixed classifier to improve the task relevance of the generator. The CondDiffEEG generator employs gated emotion-subject-time conditioning for subject-calibrated synthesis, while orthogonality regularization discourages linear overlap between the projected emotion and subject conditions. Temporal and frequency-domain constraints further preserve essential signal characteristics. For classification, CGODE-CCT first factorizes temporal and inter-channel dependencies through criss-cross attention, then models continuously evolving electrode interactions using a dynamically updated graph ODE, and finally refines the graph-enhanced representations for emotion prediction. Experiments on DEAP, SEED, and DREAMER under subject-independent and subject-dependent protocols demonstrate consistent improvements over conventional classifiers, EEG Transformer models, and generative augmentation baselines. Quantitative evaluation further indicates improved spectral consistency, real-synthetic distributional similarity, downstream transferability, and sample diversity. These findings demonstrate that task-oriented coordination between conditional EEG generation and continuous graph-based classification provides an effective approach to robust EEG emotion recognition.