GraphBet: Prospective FDR-Controlled Discovery from Temporal Graph Predictions

Temporal graph models rank likely future interactions, but a high score does not establish that a nominated group attracts more subsequent activity than its available opportunities explain. We introduce GRAPHBET, which freezes a nominated destination set and compounds predictable likelihood ratios from later interactions relative to pre-event opportunity probabilities. Under the directional null that the set attracts no more probability than those opportunities permit, the resulting wealth is an e-process. We prove graph-wide validity for adaptively selected, dependent candidates with history-dependent null labels; e-value Benjamini–Hochberg (e-BH) controls the false discovery rate (FDR) at a common reporting time, and the accumulated union of earlier e-BH selections also controls FDR. We additionally establish safe aggregation of betting experts, robustness to bounded reference error, and finite-sample links from predictive accuracy and betting regret to detection. We observe no final-snapshot false discoveries in 300 controlled global-null streams or 200 exact-law benchmark replays. In a crossed study of three prespecified training fits and 20 shared held-out semi-synthetic Wiki replays, learned candidates raise mean conditional testing power from .335 to .418 under a common fixed bet, with positive evidence-growth gains in every fit. A prespecified six-expert mixture raises power to .452 and achieves the highest mean end-to-end recall among the evaluated pipelines, .084 versus .069 for degree–recency candidates with the fixed bet, with no final-snapshot false discoveries. Thus the opportunity law governs validity, candidate learning governs what is nominated, and betting calibration governs how efficiently evidence accumulates.