ESCORT Fine Grained Urban Crime Risk Inference Leveraging Heterogeneous Open Data

Crime has been a complex social problem worldwide, impacting numerous individuals in both property and psychology, and affecting public safety as well. To prevent and avoid crime is of great importance for urban authorities and citizens. Crime prediction using various urban sensing data provides a promising paradigm to cope with this challenging problem. However, previous studies on crime prediction mainly focus on coarse-grained prediction at the region level, which fail in fine-grained road-level crime inference due to data sparsity problems. This article proposes a road-level framework for fine-grained crime risk inference leveraging heterogeneous open data. We first extract a set of features from relevant external datasets and historical crime records, i.e., temporal, spatial, and recurrence crime features. We then establish a spatio-temporal crime pattern using spatio-temporal features, and further infer road-level crime risk incorporating with recurrence crime features. The framework not only considers a useful spatio-temporal pattern but also considers a unique crime near-repeat phenomenon, which helps address the data sparsity. We further present an application crime risk-aware route recommendation to demonstrate the effectiveness of our framework. This application cannot be solved well with coarse-grained (e.g., region-level) crime risk results since it is inconvenient for people to travel while avoiding large regions with potential crime risks. Experiments with the real-world datasets from New York City show that our framework accurately infers road-level crime risk and outperforms other baseline methods.