Incorporating park events into crime hotspot prediction on street networks: A spatiotemporal graph learning approach

Abstract

Park events elevate crime risk in and around parks for brief periods by granting offenders close contact with abundant suitable targets in outdoor spaces. This study proposes to capture the formulated transient crime risk with a network-based feature, Park Event Density (PED), that monitors the dynamic event density across parks. We incorporate the PED into various crime hotspot prediction models to test its effectiveness. The sample includes all the robbery(n = 1555) and theft(n = 22596) incidents between 2016 and 2018 in the Center Side of Chicago. We generate daily and intraday crime hotspot predictions using two spatiotemporal graph learning algorithms (i.e., Graph Wavenet and Spatiotemporal Graph Convolution Neural Networks) and a traditional counterpart (i.e., LSTM). The results reveal that the PED-incorporated models have up to 25% higher accuracy, particularly in the intraday theft predictions. Another significant result indicates that the predictive accuracies of spatiotemporal graph learning algorithms are up to three times higher than their traditional counterpart. The proposed method provides additional information to security decision-makers with crime hotspot prediction models sensitive to the changing crime risk landscape across a region during park events. It also helps organize safer outdoor public events by enacting timely security interventions through more accurate crime hotspot predictions.