Variable selection for hedonic model using machine learning approaches: A case study in Onondaga County, NY

Based on the theoretical foundation of hedonic methods, positive relationships between various types of environmental amenities and house sales price have been investigated. However, as hedonic theory does not provide any arguments in favor of specific sets of independent variables, this lack of theoretical support led researchers to select independent variables from empirical results and intuitive information of previous studies. In previous hedonic studies, the most widely used selection criterion was stepwise selection for multiple regression with ordinary least square (OLS) regression for model fitting. The objective of this study is to apply machine learning approaches to the hedonic variable selection and house sales price modeling. Two rule-based machine learning regression methods including Cubist and Random Forest (RF) were compared with the traditional OLS regression for hedonic modeling. Each regression method was applied to analyze 4469 house transaction data from Onondaga County, NY (USA) with two different neighborhood configurations (i.e., 100 m and 1 km radius buffers). Results showed that the RF resulted in the highest accuracy in terms of hedonic price modeling followed by Cubist and the traditional OLS method. Each regression method selected different sets of environmental variables for different neighborhood. Since the variables selected by RF method led to make an in-depth hypothesis reflecting the preferences of house buyers, RF may prove to be useful for important variable selection for the hedonic price equation as well as enhancing model performance.close4