Machine Learning for House Price Prediction

Abstract

Housing stands as a core human necessity, and residential property prices—shaped by elements such as local crime levels and air quality—hold great significance for all stakeholders in the real estate sector. A dependable price prediction system plays a key role in supporting well-informed decisions related to property transactions. This research utilized the Boston housing dataset sourced from Kaggle and carried out all-round data preprocessing: it addressed missing or incorrect values, removed outlier data points, applied Z-score normalization to reduce data complexity, and conducted data transformation as part of the data wrangling process. Three machine learning models were evaluated in the study, namely Support Vector Regression (SVR) with a radial basis function kernel, a five-layer Artificial Neural Network (ANN) consisting of three dense layers and two dropout layers, and Extreme Gradient Boosting (XGBoost). Additionally, five-fold cross-validation was employed to select regularization parameters, and principal component analysis was used to enhance model performance. Experimental results revealed that the Artificial Neural Network outperformed the other two models, achieving the highest R-squared value of 0.86, the lowest Mean Squared Error of 0.0046, and the smallest Mean Absolute Error of 0.047. Support Vector Regression also delivered strong performance, with an R-squared of 0.83 and a Mean Squared Error of 0.0054, while XGBoost exhibited relatively higher errors, with a Mean Squared Error of 0.1214.