Predictive policing hits a speed bump

predictive-policingAs big data has become more widespread, a growing number of areas have attempted to improve their predictive capabilities, whether in medicine or investing.  One of the more interesting applications however has been in policing.

In 2014 I wrote about the ATHENA project that is being run by West Yorkshire police.

The project is aiming to bring the public into policing a lot more than is currently the case by using social and mobile tools as a force for good.  The project team concede that the public are usually first on the scene of an incident, and are often therefore well placed to relay crucial information to officials.

Taking this a step further was a project run by UCLA alongside the LAPD to use big data to position officers where they thought crimes would take place.  The reduction in crime that resulted from the project led some to herald a new dawn of predictive policing.

A bump in the road

The movement has not been without some somewhat significant hitches however.  There has been ongoing concern about the bias attached to certain algorithms, with one used in parole hearings discriminating against black offenders.

Likewise, a recent report from the RAND Corporation reveals that an experiment by the Chicago Police Department was also less than successful.

The project, which aimed to reduce gun crimes, used models to trawl through arrest data and provide a list of people that were believed to be at high risk of gun crime.  The hope was that this data could then be used to keep them from harm.

Alas, things didn’t really turn out like that, with the report highlighting two main issues with the approach.  Firstly, the team struggled to get officers to actually use the data, with around 2/3 of cases seeing the list produced by the model ignored.

This was often because officers felt ill equipped to deal with the information available to them, with no apparent direction given as to how to manage those individuals contained on the list.  Rather than being empowered by the data therefore, the officers often felt confused by it.

Equally, when the officers did act upon the information, the results weren’t particularly good, with just nine arrests made, even though the data was actually quite good.  The study found that those on the list were around three times as likely to be arrested for a firearms offense as those not on it.

It’s perhaps fair to say, therefore, that there are still a number of teething problems to iron out before this kind of technology achieves the results first hoped for.  I’m sure that the lessons will be learned however and that future applications of big data in policing will become smarter and more accepted by officers.

For now though, the age of predictive policing is a little way off.

Facebooktwittergoogle_plusredditpinterestlinkedinmail

Leave a Reply

Your email address will not be published. Required fields are marked *