Before trusting a predicted edge, ask whether its shared neighbours are informative or merely high-degree hubs.

This separates genuine local similarity from a mechanical popularity effect. It also directs attention toward degree-normalised scores and appropriate baselines.