Forecasting scientific relations can steer discovery by spotting promising links before they appear. Existing methods usually model concept semantics and graph structure separately or aggregate semantics over coarse historical snapshots, which can misalign representations with rapidly changing graph evidence. We introduce a time‑aligned evolving concept graph framework that jointly captures semantic and structural evolution. The key idea treats dated papers as shared update events, reconstructing both semantic and structural states from the same publication history at each prediction time. Pair‑level fusion then combines these states to predict first co‑occurrence, relation formation, and conditional relation type. With the architecture and training fixed, merely refreshing the context alongside graph updates boosts mean relation AUPRC by 16.6% over a frozen‑context baseline. On a graph built from 187,848 papers, 270,687 concepts and 7.45 million co‑occurrence links, the full framework raises mean relation AUROC from the strongest baseline’s 0.9290 to 0.9722, and achieves a population‑weighted mean AUPRC of 0.005778.
Review