Explainable AI for Crime Prediction in São Paulo: Addressing Underreporting and Crimes of Opportunity - CANCELLED

• In Event: Computational Methods in Criminal Justice Settings

Abstract

Predicting street crimes is a challenging task due to the complex nature of the data: it involves time series, a city street graph structure, and a sparse, imbalanced distribution of events. Additionally, crimes are often underreported, and crimes of opportunity—those occurring spontaneously without a clear pattern—further complicate predictions. To address these challenges, we propose a model combining Transformer networks, graph convolutions, and kernel density estimation. Our study focuses on São Paulo, Brazil, the largest city in Latin America, known for its sharp social inequalities and varying crime rates across different regions. We analyze multiple types of street crimes using diverse data sources, including public infrastructure, services, and weather conditions. To mitigate data sparsity, we incorporate spatial information and apply smoothing techniques to the crime time series. Despite using state-of-the-art methods, underreporting and crimes of opportunity may significantly impact model performance. To explore this, we introduce an alternative application of Explainable Artificial Intelligence (XAI) to identify potential cases of underreporting and opportunistic crimes. This, along with insights into recurring crime patterns, could contribute to more effective public security policies.

Authors

• Victor Hugo Barella, University of São Paulo

• Priscylla Maria Da Silva Sousa, Universidade de São Paulo

• Luis Gustavo Nonato, University of São Paulo