Disentangling Dynamics: A Compositional Framework for Temporal Graph Foundation Models

Learning general representations for temporal graphs is challenging because temporal dynamics and structural changes are often intertwined, causing models to capture domain-specific patterns over generalizable ones. This is acting as barrier to the development of foundation models for temporal graphs. Current methods use sequential models alongside Graph Neural Networks (GNNs), producing entangled spatio-temporal representations that struggle to transfer to new graphs. In this work, we highlight that this limitation is structural rather than statistical. To address this, we propose TIDES, a compositional framework that systematically decouples time from space during learning, serving as a stepping stone toward foundation models for temporal graphs. TIDES relies on a one-time, parallel extraction process to create three domain-agnostic priors: (i) temporal embeddings from a frozen time-series foundation model; (ii) deterministic summaries of local temporal contexts; and (iii) supra-Laplacian positional encodings of the changing graph structure. These features are computed once and used as node inputs for a GNN. TIDES achieves strong performance on multiple benchmarks, reduces negative transfer, and improves cross-graph transferability, outperforming state-of-the-art temporal graph baselines on 20 unseen networks in the MiNT zero-shot benchmark. Overall, TIDES offers a systematic approach to Temporal Graph Foundation Models by providing scalable, transferable representation strategies.