Apartment valuation models for a big city using selected spatial attributes

Abstract

This paper addresses a property valuation problem with machine learning models with pre-selection of attributes. The study aimed to examine to what extent the environmental attributes influenced real estate prices. Real-world data about purchase and sale transactions derived from a cadastral system and registry of real estate transactions in one of Polish big cities were employed in the experiments. Machine learning models were built using basic attributes of apartments and environmental ones taken from cadastral maps. Five market segmentations were made including administrative cadastral regions of a city and quality zones delineated by an expert, and classes of apartments. Feature selection was accomplished and property valuation models were built for each division of a city area. The study allowed also for a comparative analysis of performance of ensemble learning techniques applied to construct predictive models.