USI - Email

Archive / INF Seminars / INF_2025_05_20_LucaPersia

USI - Email

	Modeling Crime Dynamics in Switzerland using Machine Learning

Host: Prof. Deborah Sulem

Tuesday

20.05

USI Campus EST, Room D5.01

12:30 - 13:30


	Luca Persia ZHAW

Abstract: Machine Learning has gained traction for its ability to process large datasets and identify (non-linear) patterns in crime data. In this talk, I will present a step-by-step machine learning approach developed to predict burglaries in the Swiss cantons of St. Gallen and Zug (Quantifying Illegal Activity: Estimating Dark Rates and Predicting Offenses | ZHAW Zurich University of Applied Sciences), gradually increasing model complexity to better capture the structure of burglary risk. We begin with a baseline classifier and explore the near repeat effect to assess spatio-temporal autocorrelation in event patterns. Next, we benchmark ensemble methods with random undersampling, following the approach of Kadar et al. (2019), and analyze the trade-off between hit rate and false positives at different probability thresholds. To better handle imbalance without distorting the data, we explore cost-sensitive loss functions, with a focus on Focal Loss (Lin et al., 2017), designed to emphasize hard, misclassified examples. We show that traditional machine learning models, even through resampling and handling of the loss function, struggle with strong data imbalance and lack of meaningful patterns in low-crime regions. This provides a basis for moving beyond discrete grid classification and considering crime as a point process in space-time.

Biography: Luca Persia is a senior researcher in applied econometrics and statistics at the Zurich University of Applied Sciences (ZHAW), where he teaches time-series econometrics at the undergraduate level. He has a MSc from the University of Trento in Finance and a MSc from Politecnico di Milano in Quantitative Finance. During his studies, his research interests focused on reinforcement learning for option pricing, Poisson jump processes, and stress testing. In 2024, he joined the Innosuisse project Quantifying Illegal Activity: Estimating Dark Rates and Predicting Offenses | ZHAW Zurich University of Applied Sciences to work on crime prediction models. Previously, he has been working in the European Central Bank as a financial data scientist, conducting econometric analyses to assess market risk for the European Banking Authority Stress Test 2023.

Luca will start in September 2025 as a Ph.D. student in Computer Science in Lugano (Università della Svizzera Italiana), supervised by Deborah Sulem, focusing on spatial statistics and crime modelling through self-exciting spatio-temporal point processes. His academic page is Luca Persia | ZHAW Zurich University of Applied Sciences.