Discovering Predictive Event Sequences in Criminal Careers

In this work, we consider the problem of predicting criminal behavior, and propose a method for discovering predictive patterns in criminal histories. Quantitative criminal career analysis typically involves clustering individuals according to frequency of a particular event type over time, using cluster membership as a basis for comparison. We demonstrate the effectiveness of hazard pattern mining for the discovery of relationships between different types of events that may occur in criminal careers. Hazard pattern mining is an extension of event sequence mining, with the additional restriction that each event in the pattern is the first subsequent event of the specified type. This restriction facilitates application of established time based measures such as those used in survival analysis. We evaluate hazard patterns using a relative risk model and an accelerated failure time model. The results show that hazard patterns can reliably capture unexpected relationships between events of different types