Automated machine learning for residential property valuation

Abstract

Many studies have demonstrated that machine learning (ML) models surpass other approaches in residential property valuation. However, developing well-performing ML models requires the deep involvement of domain experts and data scientists, posing challenges for the scale of artificial intelligence (AI) implementation and application. Automated ML (AutoML) implements the end-to-end process of repetitive ML tasks without much human assistance. Although existing AutoML techniques have yielded promising results, no domain-specific AutoML framework is designed for residential property valuation. Most AutoML frameworks are domain-agnostic, overlooking domain knowledge, which limits their model performance and the practical applications of AI in real-world scenarios. This study proposes a domain-specific AutoML framework for residential property valuation (AutoML4RPV), which incorporates both domain-specific and domain-agnostic function modules to streamline AI implementation. Three real housing transaction datasets (New York, London, and Singapore) and a public valuation dataset are used for experimental validation by comparing AutoML4RPV with benchmark ML models and openly available AutoML frameworks. Integrating domain-specific modules, AutoML4RPV achieves the best performance with raw data input, outperforming the second-ranked domain-agnostic AutoML framework by 17.6 %, 18.2 %, and 9.80 % for the New York, London, and Singapore datasets, respectively. The domain-agnostic modules of AutoML4RPV are highly competitive and achieve the best model performance for the public valuation dataset. The data and codes of AutoML4RPV are openly available at https://github.com/Linhkust/AutoML4RPV.