The Impact of Feature Engineering on House Price Prediction

Details

Supervisors

Pecheur, Charles

Faculty

Ecole polytechnique de Louvain

Degree label

Master [60] en sciences informatiques

Abstract

Predicting the sale price of a house is a supervised regression problem on structured, tabular data, where the outcome depends on many property attributes at once. Most existing work on this task focuses on choosing and tuning the model, treating the preprocessing that comes before it as a routine step. This thesis takes the opposite view. Using the Ames housing dataset from the Kaggle competition House Prices: Advanced Regression Techniques, we fix the model and instead measure how much the preprocessing pipeline affects performance. We first confirm that XGBoost outperforms Random Forest on this data under identical conditions, and then run a systematic sweep over missing-data strategies and feature selection methods, evaluating every combination with five-fold cross-validation. The central finding is that the choice of feature selection method has a much larger effect on predictive accuracy than the choice of missing-data strategy: recursive feature elimination performs best, while changing the imputation strategy barely moves the error, confirming that XGBoost is largely insensitive to how missing numerical values are filled in. The final model reaches a cross-validated RMSLE of about 0.12, accurate for the majority of houses but consistently under-predicting the most expensive ones, for which the data offers too few examples.