Além dos números : uma perspectiva multimodal na avaliação imobiliária

Abstract

Real estate valuation is a classical research problem that spans several fields of knowledge. In Computer Science, this topic has been mainly addressed through the proposition of Automated Valuation Models (AVM), which aim to automate this process. However, most of these models base their estimates solely on the structural characteristics and geographical location of the properties, neglecting crucial factors such as condition and solar orientation. To support decision-making, an AVM needs to evaluate the full spectrum of information considered by a human appraiser, including photos and text. In this context, this work presents a multimodal architecture of interchangeable components that facilitates the rapid prototyping of multimodal models. This architecture was applied to a case study in Brazil, a country whose urban landscape is still underexplored in multimodal real estate valuation literature. Using data from 212,076 listings and 1,702,960 images, the observed results emphasize the relevance of image and text modalities in optimizing AVM performance. The use of textual descriptions, in particular, led to significant reductions in RMSE, ranging from 7 to 27% compared to two different baseline models.