Temporal Sheaf Neural Networks with Dynamic Orthogonal Transport

Temporal Sheaf Neural Networks (TSNN) have been introduced as a novel framework for temporal link prediction, utilizing time-varying orthogonal frames to model node-specific interactions. The architecture employs low-rank Householder products for parameterization, ensuring exact preservation of hidden states during frame updates, and incorporates a geometric-residual decoder for accurate predictions based on transported distances. TSNN demonstrates superior performance on link-prediction benchmarks, particularly in scenarios with significant node-role heterogeneity, highlighting its potential for more nuanced modeling of dynamic graph structures in AI applications.