Estimation of missing prices in real-estate market agent-based simulations with machine learning and dimensionality reduction methods

Abstract

The opacity of real-estate market involves some challenges in their agent-based simulation. While some real-estate Web sites provide the prices of a great amount of houses publicly, the prices of the rest are not available. The estimation of these prices is necessary for simulating their evolution from a complete initial set of houses. Additionally, this estimation could also be useful for other purposes such as appraising houses, letting buyers know which are the best offered prices (i.e., the lowest ones compared to the appraisals) and recommending the buyers to set an initial price. This work proposes combining dimensionality reduction methods with machine learning techniques to obtain the estimated prices. In particular, this work analyzes the use of nonnegative factorization, recursive feature elimination and feature selection with a variance threshold, as dimensionality reduction methods. It compares the application of linear regression, support vector regression, the k-nearest neighbors and a multilayer perceptron neural network, as machine learning techniques. This work has applied a tenfold cross-validation for comparing the estimations and errors and assessing the improvement over a basic estimator commonly used in the beginning of simulations. The developed software and the used dataset are freely available from a data research repository for the sake of reproducibility and the support to other researchers.