EEG-based emotion recognition is a pivotal enabler of affective brain–computer interfaces, yet existing approaches remain limited by biologically implausible modeling assumptions that undermine both accuracy and generalizability. Current graph and attention-based frameworks typically rely on static adjacency matrices or heuristic spatial priors, disregarding the dynamic and biophysically constrained nature of neural activity. To overcome these barriers, we present the EEG graph attention network (EGANet)—the first architecture to explicitly embed neurophysiological principles into an adaptive graph attention framework. EGANet introduces a spatial decay matrix derived from 3D Talairach coordinates to model inverse-square cortical potential attenuation, dynamically integrating physical distance constraints with learnable emotion-specific couplings. Complemented by rotary positional embeddings and a Hadamard product–based aggregation scheme, EGANet preserves electrode topology while capturing long-range spatiotemporal dependencies. Comprehensive evaluations on the DEAP, SEED, and FACED benchmarks demonstrate that EGANet establishes a new state of the art in EEG-based emotion recognition, delivering consistent gains in both accuracy and cross-dataset generalizability. These results highlight the transformative potential of neurophysiology-guided graph models in advancing robust, biologically grounded affective computing.