How It Works
PredPol is a predictive modeling technology built for law enforcement agencies in the US. The model aims to predict the next time and place where criminal activity may occur. It requires large swaths of crime data in order to produce reliable analytics. PredPol is similar in form and function to seismic technology, which offer scientists the probable locations for upcoming earthquake aftershocks. The primary task of the model is to translate basic crime statistics into a geographical visualization. This sets it apart from other widely-used recidivist risk models, many of which attempt to predict the individual who may commit a crime. The project began in 2012 as a collaborative experiment between the Los Angeles Police Department and UCLA professor Jeffrey Bratingham.
Input Data
PredPol technology requires at least three inputs to construct an analytical prediction and geographical visualization, all of which relate to past criminal activity. They include crime type, location, and date/time.
Latent Variables and Bias
A decade after the launch of PredPol, a study found that the model disproportionately highlighted low-income, non-White communities. The technology has no inherent biases hidden in it's framework; it is simply an analytical tool that relies historical data. This brought two issues to the forefront: the accuracy of the model, and how law enforcement agencies choose to respond to the model's predictions.
Concluding Remarks
PredPol's co-founders conducted a study in 2018 that found, had the technology been used in Indianapolis, police patrols would potentially target Hispanic/Latino communities at a rate 400% higher than White communities. According to CEO Brian MacDonald, the report was not shared with PredPol's law enforcement customers since it "was an academic study conducted independently of PredPol." The research was presented at the 2018 IEEE International Conference. The authors of the study devised a potential algorithm change which they said produced a more evenly distributed distribution of crime forecasts. However, they discovered that its forecasts have been less precise than the existing algorithm, albeit it was still "possibly more accurate" than human estimates. MacDonald made clear that the PredPol's software was not altered as a result of the study.
Reference
Dhruv Mehrotra