Interpretable Decision Trees for Place-Based Urban Crime Analysis: Socio-Spatial Patterns of Theft and Robbery in Itaquera, São Paulo, Brazil
Year: 2026
Type: article
Abstract: We examined how residential characteristics and tract size relate to recorded theft and robbery in Itaquera, São Paulo. The analytical sample comprised 39,871 police records from 2017–2021 in 703 census tracts. We fitted separate C4.5 decision trees for theft, robbery, and their combined count using five 2010 census indicators. With these predictors, mean income was the first split for every outcome. Accuracy on geographically held-out tracts was 62.2% for theft, 63.3% for robbery, and 64.2% for combined incidents. When we added tract area and resident population, area replaced income at the root, while the accuracy rose to 73.0%, 76.1%, and 78.1%, respectively. Adding predominant commercial land use did not improve these results. Recorded counts exhibited positive spatial autocorrelation (global Moran’s I, 0.383–0.395). Between 2017–2019 and 2020–2021, tract rankings remained strongly correlated and income remained the first split in the five-predictor models, but few complete rules recurred. The trees identify historical, place-based patterns in recorded counts. Their dependence on tract area and the instability of small terminal branches limit their use for inference about individual risk or future incidents.
Source: Urban Science
Institution: Universidade Nove de Julho
Language: English
Subfield: Sociology and Political Science
Field: Social Sciences
Domain: Social Sciences
Open Access status: gold
Estimated APC paid: $1,930
OpenAlex ID: W7234772136
