Student dropout prediction through machine learning optimization: insights from moodle log data

Area of Science:

• Educational Technology
• Artificial Intelligence
• Data Science

Background:

• Student attrition and academic failure are significant educational challenges requiring early identification and intervention.
• Learning Management Systems (LMS) like Moodle generate rich datasets suitable for predictive analytics.
• Existing methods struggle with timely identification due to data limitations and class imbalance.

Purpose of the Study:

• To advance dropout and failure prediction using machine learning on Moodle student activity logs.
• To investigate the effectiveness of the CatBoost algorithm for identifying at-risk students.
• To address challenges of limited and imbalanced datasets through advanced techniques.

Main Methods:

• Employed the CatBoost algorithm trained on Moodle student activity logs.
• Utilized Adaptive Synthetic Sampling for data balancing.
• Applied Non-dominated Sorting Genetic Algorithm II for multi-objective hyperparameter optimization.
• Compared models trained on weekly data versus a single model trained on all data.

Main Results:

• The model trained on all weeks' data significantly outperformed models trained on weekly data.
• Demonstrated substantial improvements in F1-scores and recall, especially for the minority class of at-risk students.
• Achieved an average F1-score of approximately 0.8 on the holdout test for the combined data model.

Conclusions:

• Machine learning, particularly the CatBoost algorithm, shows strong potential for early identification of at-risk students.
• Targeted ML approaches can facilitate timely interventions, leading to improved educational outcomes.
• Integrating LMS data with advanced ML techniques offers a promising avenue for addressing student attrition.