Explaining Neural Networks in Property Valuation

Abstract

Automation is an important aspect of mass appraisal because of quality and time requirements. Regression analysis has been used in computer-assisted mass appraisal (CAMA) for a long time but suffer of problems, like functional form miss-specification and multicollinearity, since real estate markets potentially have nonlinear effects and inter-related variables. In the last years, several papers presented artificial neural networks (ANNs) like an alternative tool to regression, with good results mainly when large samples were used. But ANN models have problems in the explanation of the results, because of its “black box nature”. This is a serious problem in fields like taxation since the taxpayer needs to be informed about the way of assessment. However, there are methods able to explain ANNs results by extracting the knowledge stored in the network parameters. Fuzzy rules extracting is the most commonly referred method, which the rules are generated from network’s weights. This paper presents a case study comparing the results of neuro-fuzzy and regression models using data of real estate market of Porto Alegre, a Brazilian city, and discussing their advantages and disadvantages.