Geospatial Machine Learning for Real Estate Valuation

The widespread digitalization of real estate transaction records, coupled with the growth in geospatial data availability and machine learning techniques, has created unprecedented opportunities for improving real estate valuation. However, effectively applying machine learning in this domain remains challenging due to the complex, spatially dependent nature of real estate data, and the need for interpretability, reliability, and regulatory compliance. This dissertation addresses these challenges by demonstrating how geospatial machine learning, combined with techniques for explainability and uncertainty estimation, can enhance the accuracy and reliability of automated real estate valuation. The research is situated at the intersection of real estate, geospatial information, and machine learning, and presents contributions that are both methodological and applied, with relevance to stakeholders ranging from financial institutions and governments to individual citizens. Part I introduces the research domain, outlining the key challenges in real estate valuation and explaining the role of geospatial information and machine learning in addressing them. It also includes a systematic literature review identifying opportunities for improved modeling through advanced techniques and richer datasets. Part II focuses on the development of tree-based models tailored for spatial data. Chapter 3 introduces GeoTree, a novel geospatial regression tree algorithm with more intuitive and spatially meaningful split types. Chapter 4 expands this into geoRF, an ensemble outperforming existing spatial trees and competing with state-of-the-art gradient boosting. Chapter 5 explores incorporating spatial penalties into gradient boosted trees to better model spatial dependence. Part III investigates deep learning methods, particularly graph neural networks, for real estate valuation. It highlights how graphs can explicitly represent the complex spatial relationships between properties, overlooked by traditional tabular approaches. The chapter evaluates various graph construction methods and graph neural network architectures, demonstrating their potential while noting graph neural networks have yet to consistently outperform gradient boosted trees in this domain. Part IV explores emerging frontiers in valuation, focusing on uncertainty quantification and large language models. Chapter 7 proposes a novel framework for uncertainty evaluation and recommends effective techniques. Chapter 8 evaluates large language models' use in low-data valuation settings, considering accuracy, explainability, and uncertainty for users lacking large datasets or professional automated valuation models. Part V concludes the dissertation by summarizing the main contributions, discussing practical implications, acknowledging current limitations, and outlining promising directions for future research.