Forecasting a changing network by predicting its shape first

A network forecast can fail before it predicts a single connection: it may start with the wrong participants. People enter online communities, transaction addresses become active, and previously active participants disappear or return. Predicting connections among a largely fixed cast does not necessarily predict the next network.

Ronan Buck and colleagues propose TOPOGED, a system that reverses the usual emphasis. Rather than starting by scoring individual future connections, it forecasts a compact structural description and then builds a graph guided by that prediction. The work is available as an arXiv preprint; its reported results should be treated as research findings rather than established deployment performance.

Turning a network into counts

A graph represents participants as nodes and interactions as edges. TOPOGED processes a sequence of snapshots, each describing the network during one time interval.

Its structural summary uses node degree: the number of connections a node has. At progressively higher degree thresholds, the encoder includes qualifying nodes and counts both those nodes and the edges between them. Early thresholds show the portion made up of less-connected participants; later thresholds admit more highly connected ones. The final counts recover the snapshot’s total nodes and edges.

This nested sequence is called a filtration. Despite the topological terminology, the method does not calculate the full persistent-homology machinery used in some other approaches. It uses inexpensive node and edge counts across degree thresholds, with ten thresholds by default.

A lightweight temporal predictor forecasts how these sequences will change. The implementation uses an exponentially weighted moving-average approach selected from five forecasting methods. It also predicts the fraction of arriving nodes and the proportions of different connection types.

Building one possible future

The predicted totals set node and edge budgets. A memory module selects returning participants from recent snapshots, favoring nodes that appeared recently, appeared repeatedly, and had higher connectivity. Selection is weighted sampling, not simply a list of the highest-scoring nodes.

New participants receive fresh identifiers. Crucially, the model predicts how many will arrive and their structural roles, not their external identities. It cannot name a previously unseen person or transaction address before observation.

The decoder then adds connections in four stages: recurring connections between known nodes, previously unseen connections between known nodes, connections joining known and new nodes, and connections between newcomers. A graph neural network supplies representations for scoring candidate pairs. Predicted budgets and soft structural guidance steer construction.

The result is one plausible snapshot, not a uniquely recovered future. Different graphs can share the same filtration counts, so matching the summary does not guarantee matching individual interactions.

What the experiments found

The researchers tested 14 datasets: messaging, MathOverflow, Reddit and Wikipedia interaction networks, plus ten Ethereum token transaction networks. They used chronological training, validation and test partitions of 70%, 15% and 15%, with evaluations over three independent runs.

Table 3 reports a macro-average node F1 score of 0.49 for TOPOGED, compared with 0.05 for the strongest baseline. F1 balances precision and recall; these values are not percentages of predictions that were correct. The reported macro-average edge F1 was 0.12, while baseline results rounded to approximately zero. The authors also report winning about 60% of graph-structure metric evaluations.

Those improvements coexist with substantial difficulty. TOPOGED’s edge F1 was only 0.02 on MathOverflow and 0.03 on Reddit. Capturing broad network structure is clearly different from identifying its exact future connections.

What the comparison establishes—and does not

The baselines were primarily link-prediction systems adapted to construct complete graphs. Their connection scores were converted into edges using validation-tuned thresholds, with a cap on graph growth. TOPOGED instead explicitly controls participant counts and connection budgets. The comparison therefore tests a demanding reconstruction setting, not ordinary link-ranking performance.

The aggregate node scores also require care: the paper describes identity-based recovery metrics while explicitly declining to predict newcomers’ external identities. They should not be read as evidence that it identifies unseen participants in advance.

Finally, ten datasets come from token transactions, and average snapshots range from dozens to roughly 1,450 nodes. These experiments do not establish performance on every network type or at much larger scales. The central contribution is narrower but useful: treating participant turnover and global structure as explicit forecasting targets can produce better future snapshots than relying on connection scores alone.