Reliability-Aware Hybrid Multi-Graph Imputation of Bilateral IRI for System-Level Pavement Monitoring

Missing international roughness index (IRI) measurements can distort pavement-condition records and maintenance priorities. This study develops RA-MGIN-H, a hybrid imputation framework that allocates estimator authority using observable relational support and evaluates the resulting information through uncertainty, downstream distortion, and review escalation. Its relational component combines route and geographic attention with a tabular/prior expert. A bounded support gate blends this estimate with CatBoostJointAug. The primary analysis uses 3,751,807 unique 2023 pavement nodes across 51 routes, with survey-year filtering before aggregation and graph construction. Across 30 controlled scenario-seed cells, RA-MGIN-H achieved a cell-weighted mean RMSE of 1.179 ± 0.118 m/km, compared with 1.322 and 1.332 m/km for separately tuned fixed dual-graph and single-graph GraphSAGE controls. Gains over the relational and boosted-tree comparators were moderate and not significant in multiplicity-adjusted seed-level comparisons. A retrospective T22-to-T23 test on seven corridors used separately constructed yearly graphs and T22-only run-level selection. Hybrid RMSE was 1.045 m/km versus 1.192 m/km for CatBoost, with no overall improvement over the relational component. Nominal-50% masking retained high aggregate route correlation, but masked-entry and spatial-block diagnostics exposed priority error. Five overlapping sensitivity replays matched to the observed field-gap profile produced hybrid RMSE of 5.453 m/km and negative mean R2 despite high relational support. The transferable contribution is the separation of evidence availability, bounded estimator authority, and downstream assurance. Controlled-mask comparisons support comparative predictive performance, whereas the replay places genuine field gaps outside the current operating envelope and requires remeasurement or manual recovery.