Detecting Key Offenders from Crime Incidents via Attributed Heterogeneous Network Learning

Targeting key offenders as well as understanding their delinquent behaviors are critical to concentrate limited enforcement resources and even to crush crime chains. Despite its importance, previous studies have not treated the key offenders identification problem in much detail. In this paper, we attempt to tackle this interesting yet open issue. First, we propose to use the Attributed HEterogeneous Network (AHEN) to model the linkage among criminals and crime incidents. Second, we define a number of features associated with both offenders and incidents. Third, we present a representation learning method working on the crime AHEN to obtain embeddings of nodes and edges. Finally, we utilize these embedding vectors as features of several typical classifiers to detect key offenders. Experimental results on real-life dataset demonstrate that with the hand-crafted features and the learned embeddings, the accuracy of detecting key offenders can be improved remarkably