Machine Learning Analysis of Student Performance

Abstract

This thesis demonstrates a systematic application of supervised machine learning techniques for predicting student's mathematics performance using demographic and academic features. Multiple regression models including Linear Regression, Ridge Regression, Lasso Regression, Random Forest, and XGBoost were developed and evaluated within a unified preprocessing pipeline to ensure a fair comparison. The results indicate that regularised linear models perform comparably to, and in some cases outperform, more complex ensemble methods, suggesting predominantly linear relationships within the dataset. Lasso Regression was selected as the final model due to its strong predictive performance and inherent feature selection capability. Further analysis identified reading and writing scores as the most influential predictors of mathematics performance. The final model was implemented in a Streamlit-based application to demonstrate its practical applicability.