Uso de aprendizado de máquina na modelagem preditiva dos valores de imóveis na cidade de João Pessoa

Abstract

Brazil’s urban trajectory underwent significant changes throughout the 20th century. In 1940, only 31% of the population lived in urban areas, but this figure had already surpassed 50% by 1970 (WAGNER; WARD, 1980). The accelerated growth of cities increased the demand for housing, leading to the emergence of public policies and financing mechanisms such as the Sistema Financeiro de Habitação (SFH) and, later, the Sistema de Financiamento Imobiliário (SFI). This process directly influenced the development of the real estate market in various cities, including João Pessoa, capital of the state of Paraíba, which has been gaining national prominence. By November 2024, the city recorded an accumulated appreciation of 16.13% in residential properties over the previous 12 months, ranking among the capitals with the highest growth in the sector. In light of this scenario of expansion and appreciation, this study aimed to develop a machine learning-based model to predict property values in João Pessoa, using data collected through web scraping from real estate websites. Various machine learning techniques were applied to build and evaluate predictive models, with the final model based on the Gradient Boosting algorithm showing satisfactory performance, with an 𝑅2 of 85.42%, a MAPE of 18.21%, and an RMSE of 227,677.44. In addition to modeling, interpretability techniques such as SHAP, LIME, ICE, and PDP were used to assess variable importance and behavioral patterns in price estimation. The results indicated that the property area was the most influential factor in predictions, followed by the ratio between the number of bedrooms and area, number of parking spaces, geographic coordinates, number of bedrooms, and the average rental value in the neighborhood. Finally, the data and predictive model were integrated into a web application developed to assist in the property valuation process in João Pessoa. The application allows users to make predictions based on provided characteristics and visualize descriptive maps and charts, offering a data-driven approach to the local real estate market.