Graph neural networks for topological blendshape dependency mapping in explainable affective computing architectures

Abstrak

This study develops an explainable spatial-temporal graph neural network (ST-GNN) architecture grounded in anatomical muscle co-activations for real-time human cognitive and behavioral state monitoring. Departing from traditional Euclidean pixel-grid deep learning, the proposed framework models facial blendshape weight coefficients as structured nodes and edges within a non-Euclidean topological karkas, processing continuous deformations via spatial-temporal graph convolutional networks (ST-GCN). Empirical validation demonstrates that the architecture mitigates sensory data degradation under physical occlusions and off-axis head rotations, achieving an exceptional macro F1-score of 0.942 and stabilizing processing latency at an ultra-low 112.5 ms threshold. To resolve the long-standing black-box limitation of deep networks, a game-theoretic Shapley additive explanation (SHAP) protocol is intrinsically integrated to map local feature attributions directly onto the graph vertices, providing a mathematically transparent and auditable decision matrix unique for mission-critical telemetry and high-stakes safety deployment.