Evaluating ensemble models for fair and interpretable prediction in higher education using multimodal data.

Area of Science:

• Learning Analytics
• Educational Data Mining
• Higher Education Research

Background:

• Online higher education faces challenges with student attrition.
• Existing predictive models often lack comprehensive data integration and comparison with advanced techniques.
• There is a need for robust, interpretable, and fair models for early prediction of academic performance.

Purpose of the Study:

• To develop and evaluate a predictive framework for early identification of at-risk students in online engineering programs.
• To integrate diverse data sources including Moodle interactions, academic history, and demographics.
• To compare the performance of various machine learning models and identify key predictors of academic success.

Main Methods:

• Utilized data from 2,225 engineering students, incorporating Moodle logs, academic records, and demographics.
• Applied SMOTE (Synthetic Minority Over-sampling Technique) for class balancing.
• Evaluated seven base learners (including Random Forest, XGBoost, LightGBM) and a stacking ensemble using 5-fold stratified cross-validation.
• Employed SHAP (SHapley Additive exPlanations) analysis for model interpretability.

Main Results:

• LightGBM demonstrated superior performance as a base model (AUC=0.953, F1=0.950).
• A stacking ensemble did not significantly improve performance and showed instability.
• Early academic grades were identified as the most influential predictors across top models.
• The final model exhibited strong fairness across demographic groups (consistency=0.907).

Conclusions:

• The developed model effectively predicts academic performance in online engineering students.
• Early grades are critical indicators of student success, informing timely interventions.
• The study provides a state-of-the-art, interpretable, and fair model for learning analytics, aiding institutions in reducing attrition.