Temporal graph prediction learns to sketch topology first: TopoGED lifts edge metrics from near zero across 14 datasets
Synopsis
Temporal graph learning usually treats future prediction as scoring edges or node labels over a largely fixed node set, an assumption that collapses once new nodes arrive and the future graph becomes a new object in size and structure. TopoGED reframes the task as an inverse topology problem: predict a multiscale topological descriptor of the future graph, then reconstruct a future snapshot that realizes it under inductive constraints. Across 14 temporal graph datasets, its node-level macro average beats the strongest baseline, it wins most graph-structure metric evaluations, and its macro-average edge prediction metrics rise from near zero to a nonzero level.
Interpretation
The target of temporal graph prediction shifts from scoring candidate edges to generating the complete next snapshot, making node churn and structural change explicit prediction targets. Most prior methods perform discriminative inference over a largely persistent node set, focusing on edges or node labels, and that assumption breaks when new nodes and edges appear. Across 14 temporal graph datasets, TopoGED's node-level macro average exceeds the strongest baseline, it wins most of the twelve graph-structure metric evaluations, and its macro-average edge prediction metrics rise from near zero to a nonzero level.
The method first predicts a topological summary of a graph, then assembles the graph from it. The descriptor is a multiscale sequence built by filtering on node degree thresholds, recording active node and induced edge counts at each threshold, with the final threshold recovering the snapshot's node and edge budgets. A predictor outputs the next snapshot's node budget, edge budget and new-node fraction; a memory module selects reappearing old nodes from recent snapshots, new nodes are instantiated to the arrival budget, and a decoder generates edges in four stages: old–old seen, old–old unseen, old–new and new–new.
Perspective
The demonstrated scope is discrete-time snapshot forecasting: on 14 temporal interaction datasets, a sequence of snapshots up to a given time predicts the next snapshot, and structure, node and edge metrics are evaluated on the forecast graph.
Newly arriving nodes are generated only by count and structural role, without their external identities, so the model cannot predict who the new nodes are before they are observed; continuous-time event generation and identity prediction for unseen nodes are listed as future directions.
