Handling Imbalanced Ordinal Grade Prediction in Educational Analytics Using Weighted Ordinal Logistic Regression

Abstract

Imbalanced ordinal outcomes are common in educational analytics, where lower academic grades occur less frequently yet remain critical for identifying at-risk students. Standard predictive models often favor majority outcomes, reducing sensitivity to minority grade categories. This study evaluates weighted ordinal logistic regression as an interpretable approach for handling imbalanced ordinal grade prediction using student gender, age, and math anxiety as predictors. Unweighted and inverse-frequency weighted models were compared using precision, recall, F1 score, accuracy, mean absolute error (MAE), quadratic weighted kappa (QWK), and area under the curve (AUC). Weighting improved detection of minority lower-grade categories (C–E), increasing macro F1, but reduced performance for majority grades and overall discriminative ability. These findings suggest that weighted ordinal logistic regression can support educational risk identification when minority-class detection is prioritized, although predictive performance remains constrained by limited sample size and predictor strength.