Analyzing the Influence of Road Network on Vehicle Theft in São Paulo Using Graph Autoencoders

• Grant number: 25/13557-0
• Support Opportunities: Scholarships abroad - Research Internship - Master's degree
• Start date: December 01, 2025
• End date: May 31, 2026
• Field of knowledge: Physical Sciences and Mathematics - Computer Science - Computing Methodologies and Techniques
• Principal Investigator: Luis Gustavo Nonato
• Grantee: Raissa Rosa dos Santos Januário
• Supervisor: Vania Aparecida Ceccato
• Host Institution: Instituto de Ciências Matemáticas e de Computação (ICMC). Universidade de São Paulo (USP). São Carlos , SP, Brazil
• Institution abroad: KTH Royal Institute of Technology, Sweden
• Associated to the scholarship: 24/07478-8 - Exploring the Relationship between Road Network and Vehicle Theft in São Paulo, BP.MS

Abstract

In recent years, deep learning techniques have proven highly effective in identifying complex patterns in data within Euclidean domains. However, many urban phenomena, such as crime, manifest spatially along the road network, resulting in linear distributions that do not conform well to the traditional two-dimensional Euclidean space. In such cases, the urban structure can be more appropriately represented as a graph, requiring approaches capable of handling the complexity of these structures.Graph Neural Networks (GNNs) have emerged as a response to this challenge, enabling representation learning from data structured as graphs. Among GNN architectures, Graph Autoencoders (GAEs) stand out for their ability to generate latent representations that simultaneously preserve both node attributes and network topology.In this context, this project proposes the application of GAEs to the representation of the road network of São Paulo, SP, Brazil, with the objective of investigating the influence of street structure on the occurrence of vehicle thefts. Grounded in environmental criminology theories, such as Routine Activity Theory and Crime Pattern Theory, this study is based on the premise that the spatial organization of urban roads shapes daily movements and the convergence of targets, offenders, and the absence of guardianship, which are central factors in the dynamics of crime.The originality of the project lies in the combination of spatial modeling aligned with urban reality, the use of advanced machine learning techniques, and the application to a socially relevant problem. Thus, it is expected not only to contribute to methodological advances in crime studies but also to provide support for more effective public policies regarding urban safety and crime prevention.