Bayesian synthesis predicts shifting network structures with calibrated forecasts

Networks—from social graphs to biological pathways—are shaped by competing structural mechanisms like community clustering, geometric proximity, or hub-and-spoke patterns. But the dominant mechanism can shift over time, and fixed-weight models fail to adapt. In a new preprint on arXiv, Marios Papamichalis, Regina Ruane, and Theofanis Papamichalis introduce a dynamic Bayesian predictive synthesis framework that treats each mechanism as an independent forecasting agent. A synthesis layer combines their edge predictions with time-varying weights, learned from data. At each step the model produces a calibrated probabilistic forecast for the next snapshot's edges, plus interpretable inference on which mechanism is driving the prediction.

The method requires a sparse-safe parametrization to handle large networks and an identification theory that proves a single graph snapshot can recover the mechanism weights. A sharp theoretical threshold separates mechanisms that are distinguishable from those that are not, and changes in the active mechanism are tracked with an optimal per-switch cost. On real-world networks, simulations, and benchmarks, the synthesis produces accurate, calibrated forecasts and correctly identifies the leading mechanism as it shifts. The work generalizes link prediction to dynamic settings where the very rules of connection evolve over time.