Data-driven teaching quality monitoring with a transformer-GNN hybrid

Universities increasingly log dense streams of student interactions, yet most dashboards act only after learning issues have become visible.To enable earlier and cleaner detection, this study introduces a Transformer-GNN hybrid model for data-driven monitoring of teaching quality.A temporal encoder first learns long-range dependencies from sequential event data, while a graph encoder integrates cohort and course context.A reliability-aware gate and per-family calibration then refine alert stability and accountability.Across three academic terms covering six course families, the model improved the area under ROC by 2.4 points and the area under precision-recall by 5.7 points over a strong fusion baseline, reduced calibration error by 45%, and extended mean warning lead time by 1.3 days.The framework remains robust under pacing shifts and provides interpretable, actionable explanations for instructors.