Explicable Machine Learning Rental Price Prediction Using Geospatial Network Data

Abstract

Hedonic pricing models of homes focus on the explainability of the value to individual components, but they typically rely on simple analytical models with lower predictive strength than machine learning models. Here we explore a hybrid approach to leverage the power of machine learning algorithms while only relying on explanatory variables. Specifically, we are interested in only using rich geospatial data that can capture the value of neighborhood and accessibility features in a general way. First, we estimate residential demand in the Tokyo area using network diffusion from all employee locations. We then use LightGBM to compare the predictive accuracy of this estimated demand versus using station-specific categorical variables and coordinates. We find slightly better results using either station names or lon/lat; likely because they pick up additional spatial characteristics. Then we test the impact of zoning, land use, stores, vegetation, population, building structures, and station importance. We find that these features improve the accuracy of predictions for all variables sets, but they do not fully compensate for the information encapsulated in the coordinates.