Influence prediction in collaboration networks: An empirical study on arXiv

We present a case study of the Social Sphere Model (SSM), an interpretable, training-free pipeline that completes a partially observed graph with simple link-prediction scores and then selects top- k nodes by heuristic centralities. Our task is influencer identification from sampled graphs : given a subsampled snapshot of a collaboration network, recover high-influence nodes in the full graph. On the arXiv GR-QC coauthorship network we examine two observation regimes (70% and 90% of edges revealed) and two horizons. We evaluate by (i) overlap with top- k centralities on the full graph and (ii) discrepancy between influence-spread curves under simple and complex contagion. Across settings, a resource-allocation variant ( RA-2 ) pairs well with SSM in this dataset. We do not claim temporal forecasting or state-of-the-art influence-maximization; rather, we document how the choice of link-prediction score affects downstream influencer rankings in a sampled-graph pipeline, with code and scripts released for exact reproduction.