Crime Hot-Spot Modeling via Topic Modeling and Relative Density Estimation

In a typical emergency call for service event, a 911 dispatcher takes down information from the caller, dispatches an officer, and attaches available location, type, time, and text information to the call record. Eventually, the officer arrives on scene, handles the situation at hand, and then writes up a detailed report of what occurred during the incident, known as an "incident narrative". Such narratives are extremely informative and contain nuance beyond the set of time-points and event category assigned by a call dispatcher and recorded in a call for service database. However, the narratives are also highly noisy and unstructured as they are often written in haste. In this work, we study an extensive call-for-service data set from the Atlanta Police Department (APD).