<![CDATA[Purpose – This study evaluates a Temporal Representation Learning framework that addresses limitations of Learning Management System (LMS) evaluation based on aggregate activity counts by mapping chronological instructional interactions into latent representations to predict institutional Lecturer Performance Scores (LPS). Methods – Long Short-Term Memory (LSTM) and Transformer encoders modeled 144,178 activity logs from 262 instructors across 1,585 classes as event sequences. Performance was evaluated using five-fold cross-validation against aggregate-frequency baselines, including Decision Tree, Random Forest, Support Vector Regression, and XGBoost. An order-shuffling ablation assessed the contribution of chronological event order. Findings – Sequential models achieved lower prediction errors than conventional baselines. LSTM achieved a root mean square error (RMSE) of 4.35 ± 0.22 and R² of 0.89 ± 0.02, while the Transformer performed best, achieving RMSE of 4.10 ± 0.19, mean absolute error of 2.95 ± 0.15, and R² of 0.92 ± 0.01. Shuffling event order reduced Transformer performance to R² of 0.63 ± 0.04 and RMSE of 9.75 ± 0.35, supporting the predictive contribution of chronology. Exploratory K-means clustering identified three teaching activity archetypes, with a silhouette coefficient of 0.41. Research implications – Chronological interaction patterns capture predictive information overlooked by aggregate metrics, supporting the development of sequence-aware institutional analytics for lecturer performance evaluation. Originality – This study applies deep sequential representation learning to extract latent instructional activity patterns directly from uncurated LMS event streams and examines their predictive value through chronological order ablation.]]>
Copyrights © 2026
