Predicting house prices with machine learning methods

Abstract: In this study, the machine learning algorithms k-Nearest-Neighbours regression (k-NN) and Random Forest (RF) regression were used to predict house prices from a set of features in the Ames housing data set. The algorithms were selected from an assessment of previous research and the intent was to compare their relative performance at this task. Software implementations for the experiment were selected from the scikit-learn Python library and executed to calculate the error between the actual and predicted sales price using four different metrics. Hyperparameters for the algorithms used were optimally selected and the cleaned data set was split using five-fold cross-validation to reduce the risk of bias. An optimal subset of hyperparameters for the two algorithms was selected through the grid search algorithm for the best prediction. The Random Forest was found to consistently perform better than the kNN algorithm in terms of smaller errors and be better suited as a prediction model for the house price problem. With a mean absolute error of about 9 % from the mean price in the best case, the practical usefulness of the prediction is rather limited to making basic valuations.