Cross-course generalizability of SRL-aligned predictive models using digital learning traces

Summary

Drawing on self-regulated learning theory, the study tests early-warning models using activity data from three undergraduate theoretical computer science courses at two universities. Weekly indicators of learning behavior were analyzed with Elastic Net, Random Forest, and XGBoost to assess both prediction accuracy and the reliability of estimated risk.

Time management, effort regulation, and continued engagement helped identify students at risk. Random Forest performed best within the original setting, while Elastic Net transferred more reliably across settings. Accuracy and calibration weakened between institutions with different rates of at-risk students. The results suggest that early-warning models should be checked in their intended course context before being used to guide student support.