Enhancing Real Estate Valuation through Deep Modeling of Spatiotemporal Dependencies and Feature Interactions

Real estate valuation is crucial for both economic stability and societal development. Accurate property valuations are essential in guiding individual decisions and macroeconomic policies, especially in the context of rapid urbanization. Recent web platforms have transformed real estate markets by using data-driven models to predict property values. Despite these advancements, challenges remain in modeling the complex interactions between continuous and categorical factors and in addressing the spatio-temporal dependencies inherent in real estate transactions. To address these challenges, we propose TabularCNP (Tabular Conditional Neural Process), a novel framework that leverages both deep learning models on tabular data and Conditional Neural Processes (CNPs). The framework incorporates deep tabular models to handle complex feature interactions, while conditioning the target property's value on nearby recent sales to capture spatiotemporal dependencies. By transforming real estate valuation into a tabular data-driven machine learning task, we effectively consider the intricate interactions between features. Additionally, we introduce a CNP- -based event logging module to capture the influence of nearby transactions. Extensive experiments on real-world datasets validate the proposed framework, demonstrating significant improvements in prediction accuracy by effectively modeling multifactor interactions and spatiotemporal dynamics. The code is available at https://anonymous.4open.science/r/TabularCNP-EE04/.