Leveraging Textual Descriptions for House Price Valuation

Luís Fernando Bittencourt, Otávio Parraga, Duncan D. Ruiz, Isabel H. Manssour, Soraia Raupp Musse, Rodrigo C. Barros

Abstract: Real estate valuation has been vastly studied by the research community, with several articles proposing Automated Valuation Models (AVM). However, most of those models base their estimates only on geographic location and structural characteristics of the property, disregarding several factors that influence prices, such as the need for repairs and sun exposure. To support decision making, an AVM needs to “look” for the same type of information a person would when valuating a property, including photos and textual descriptions. In this work, we show that the usage of textual data can significantly increase the performance of house price-prediction models. Our experiments explore different combinations of learning algorithms and methods to extract relevant information from textual descriptions, with some surprising conclusions regarding the best combination of approaches. Overall, we shed some light on how textual features can be leveraged by the models, explaining the paths that lead to predictions that end up resulting in performance gains.