Handling Dual Skewness in Citywide Crime Prediction Using Personalized Graph Federated Learning

The widespread deployment of sensors and recent advances in machine learning have enabled the development of predictive services in smart cities, particularly in the domain of public safety through predictive policing across different regions. Recently, training on these distributed regions is facilitated by federated learning (FL) without exposing sensitive data. However, crime data exhibit inherently complex spatiotemporal dynamics and suffer from dual skewness, as both the quantity and category distribution of crimes vary significantly across regions. Most existing federated methods struggle to effectively handle such skewed distributions, while also being limited to capture spatial dependencies. To overcome these challenges, we propose an approach called g raph s tructured p ersonalized f ederated l earning ( GSpfl ), which jointly models spatial and temporal patterns in crime data. Specifically, GSpfl incorporates graph based aggregation at the server to model spatial dynamics, where similarity based adjacency matrix enable convolution among related clients to mitigate quantity skewness. On the client side, temporal dynamics are learned locally with regularization terms to balance global knowledge sharing and local model adaptation, while Focal Tversky loss is employed to address category imbalance. Experiments on real-world crime datasets from Los Angeles (LA) and Chicago (CHI) demonstrate the superior performance of our approach compared to ten baseline and state-of-the-art FL methods

Data modeling · Feature learning · Graph · Pooling · Skewness · Software deployment · Support vector machine · Crime Patterns and Interventions · Human Mobility and Location-Based Analysis · Smart Cities and Technologies