RuleCast: Evidence-Constrained Partner-Role Forecasting for Temporal Link Prediction

Existing temporal link predictors aggregate source history and score candidate compatibility, which overlooks the structural regularity of future partner roles. We propose a fundamentally different factorization: future links arise from evidence-constrained partner-role forecasting. First, the source's historical interactions predict the expected behavioral role of its next partner. Second, candidate nodes are grounded by candidate behavior and query-specific historical evidence. Our RuleCast model implements this paradigm with time-varying soft roles and a protocol encoder that forecasts role distributions. It employs specialized experts to model role compatibility, history matching, reciprocity, closure, and periodicity, alongside a candidate-aware routing mechanism that activates only valid evidence-supported experts. A lightweight identity residual preserves transductive node characteristics. Evaluated on ten temporal graph benchmarks, RuleCast consistently improves MRR and improves AP and ROC-AUC on most datasets. Ablations verify that role forecasting and evidence grounding are mutually essential, validating our new paradigm for temporal link prediction.