Using machine learning to forecast domestic homicide via police data and super learning

Abstract

We explore the feasibility of using machine learning on a police dataset to forecast domestic homicides. Existing forecasting instruments based on ordinary statistical instruments focus on non-fatal revictimization, produce outputs with limited predictive validity, or both. We implement a "super learner," a machine learning paradigm that incorporates roughly a dozen machine learning models to increase the recall and AUC of forecasting using any one model. We purposely incorporate police records only, rather than multiple data sources, to illustrate the practice utility of the super learner, as additional datasets are often unavailable due to confidentiality considerations. Using London Metropolitan Police Service data, our model outperforms all extant domestic homicide forecasting tools: the super learner detects 77.64% of homicides, with a precision score of 18.61% and a 71.04% Area Under the Curve (AUC), which, collectively and severely, are assessed as "excellent." Implications for theory, research, and practice are discussed.

Figure 1

ROC Curves. The London Metropolitan Police Service’s (MPS) standard risk assessment appears on the left, whereas the super learner appears on the right.

Figure 2

Illustration of the Super Learner. Phase One depicts a dataset being ran through various machine learning models, which each independently predict whether an individual will commit a domestic homicide. The results of these multiple predictions are stored in the Super Dataset. Phase Two depicts a separate machine learning model—the super model—generating one final homicide prediction from the super dataset—a dataset that contains the prediction of other machine learning models.