Disentangling Dynamics: A Compositional Framework for Temporal Graph Foundation Models

TIDES (Temporal Injection of Decoupled Encodings into Static GNNs) decouples time from space: three frozen, domain-agnostic priors (a time-series foundation model, local statistical summaries, and supra-Laplacian positional encodings) are computed once and fed to a static GNN, improving zero-shot transfer to unseen temporal graphs.

• Interactive demo
• Paper:

Download Paper