Modeling Student Retention and Success Using Explainable Machine Learning Techniques

Abstract

This report presents a reproducible study of student dropout and score prediction using the Open University Learning Analytics Dataset (OULAD) and the code, experiments and outputs contained in the provided Jupyter notebook. The OULAD tables were joined and student-level features were engineered by aggregating Virtual Learning Environment (VLE) interactions (total clicks, active days, distinct resources) and linking demographic and assessment records to each student instance [1]. An 80/20 stratified train–test split produced 32,640 training rows and 8,161 test rows, and preprocessing was implemented via a Column Transformer pipeline: numeric imputation and scaling plus one-hot encoding for categorical fields. To address class imbalance the training set was synthetically balanced with SMOTE producing a balanced class distribution (17,996 examples per class) prior to model training [2], using the imbalanced-learn implementation called from the notebook [4]. A consistent evaluation helper computed accuracy, weighted F1, precision, recall and produced normalized confusion matrices for each fitted model; model explanations for the Elastic Net logistic model were explored with a SHAP demo on a small test subset [5]. The notebook trains the following classifiers (as implemented with scikit-learn): Elastic Net Logistic Regression, Gaussian Naive Bayes, K-Nearest Neighbors (k=7, distance weights), SVM (RBF kernel), MLP Classifier (two hidden layers 128,64 with early stopping) and a soft voting ensemble combining selected members [3]. Reported test set performances (from the notebook runs) are: Gaussian NB — Accuracy 74.94%; KNN — 77.98%; Elastic Net Logistic Regression — 80.88%; SVM (RBF) — 83.81%; MLP Classifier — 84.93%; Soft Voting Ensemble — 84.21%. Normalized confusion matrices and classification reports for each model are displayed in the notebook and used to compare per-class recall and precision. Taken together, the notebook shows that (1) preprocessing with careful aggregation of VLE interactions plus class rebalancing materially improves classifier performance on OULAD-derived targets; (2) a tuned MLP achieved the highest single-model accuracy on the test split (≈84.9%); and (3) model interpretability was briefly demonstrated using SHAP for the logistic model to surface feature contributions on a test subset. All experiments, metrics, plots (including confusion matrices) and numeric values reported here are reproduced directly from the executed notebook cells and their outputs; external references below cite the dataset and the primary tool / method papers used or referenced in the notebook.