Multimodal temporal feature fusion for teacher competency assessment and precision training resource recommendation

Assessing teacher competency in a reliable and multidimensional manner remains an open problem, largely because conventional evaluation instruments capture only a fraction of the behavioral repertoire that defines effective instruction. We tackle this challenge by developing an integrated framework that fuses heterogeneous classroom signals-video, audio, transcribed text, and physiological recordings-through modality-specific encoders coupled with a cross-modal attention mechanism. The attention module adaptively re-weights each data stream according to its diagnostic relevance for a given competency dimension, while a hierarchical temporal component jointly models short-term pedagogical adjustments and long-term professional growth trajectories. Competency scores are formulated as a continuous regression task (evaluated via RMSE and MAE) and simultaneously discretized into ordinal proficiency levels for classification-based evaluation (accuracy and F1-score), thereby addressing both assessment perspectives within a unified multi-task objective. A knowledge graph-enhanced recommendation engine then maps diagnosed competency gaps onto targeted training resources. Experiments conducted on multimodal recordings from 856 teachers across 15 schools demonstrate that our model reaches 0.834 classification accuracy and 0.312 RMSE, outperforming all baselines on each of the seven evaluation dimensions. The recommendation module attains 0.478 Precision@5, a 13.0% relative gain over the strongest knowledge-graph baseline. Ablation analyses confirm that every architectural component contributes measurably; removing temporal modeling alone reduces accuracy by 7.1 percentage points. Taken together, these results establish a closed-loop, interpretable pipeline from diagnostic assessment to actionable professional development pathways.