Machine LearningAugust 26, 20232 min read

Student Dropout Prediction

Problem Statement

Higher education dropout is a costly problem for institutions and students alike. Using a UCI Machine Learning dataset of 4,424 students with 36 enrollment-time features, this project builds a classifier to identify at-risk students early enough for intervention.

The dataset captures academic path, demographics, and socio-economic factors known at enrollment - meaning the model could flag students before they ever attend a class.

Methodology

Preprocessing & Class Imbalance

The original three-class target (Dropout, Enrolled, Graduated) was collapsed into a binary problem: Dropout vs. Retained. The resulting class imbalance was addressed with SMOTE oversampling on the training set.

Class distribution before and after SMOTE balancing — Class distribution before and after SMOTE resampling, ensuring balanced training data.

Feature Engineering

Continuous variables were scaled using Min-Max normalization, fit on the training set and applied consistently to the test set to prevent data leakage.

Results

Model Performance

A CatBoost Classifier achieved strong performance on the held-out test set:

Accuracy	Precision	Recall
89%	87%	91%

Confusion matrix for the CatBoost model — Confusion matrix shows strong classification performance with few false negatives.

ROC and Precision-Recall curves for the model — ROC and Precision-Recall curves confirm robust discrimination between dropout and retained students.

Feature Importance

Top features driving dropout prediction — Top predictive features reveal that tuition status and curricular performance dominate the model.

Financial & Demographic Risk Factors

Students with tuition debt, without scholarships, and from lower-income backgrounds showed meaningfully higher dropout rates - suggesting targeted financial aid could be a powerful retention tool.

Dropout rate by tuition payment status — Students with outstanding tuition have significantly higher dropout rates.

Dropout rate by scholarship status — Scholarship holders are substantially more likely to complete their degree.

Dropout rate by gender — Gender-based dropout rate comparison reveals differences in retention patterns.

Next Steps

Deploy the model as a scoring API for real-time enrollment risk assessment
Apply SHAP values for per-student explainability to guide targeted interventions
Explore ensemble methods (stacking, blending) for incremental accuracy gains