London’s Future Crime Hot Spots Predicted Using Mobile Phone Data

Data mining can spot future crime scenes with an accuracy of 70 per cent, say data scientists

In the 2002 film, Minority Report, the police apprehend criminals by predicting that they are about to commit a crime. Many observers have drawn parallels between this and the modern trend towards “predictive policing”. This is the way a growing number of police forces around the world are using data on past crimes to predict the likelihood of crimes in the future.

And the results have been generally favourable, with several forces claiming that it allows them to allocate resources more effectively. For example, if there have been burglaries in your neighbourhood in the recent past, the algorithm would flag it as a potential future crime scene. And that allows law enforcement organisations to react accordingly, such as patrolling high risk areas rather than low risk ones. The idea is to stop crime before it starts.

But that raises an interesting question. How good can predictive policing become? And what kind of data will police forces use to make these improvements?

Today we get an answer of sorts thanks to the work of Andrey Bogomolov at the University of Trento in Italy and a few pals who have used the daily data from mobile phones to significantly improve the accuracy of crime predictions. “With real crime data from London we obtain an accuracy of almost 70% when predicting whether a specific area in the city will be a crime hotspot or not,” they say.

Here’s how it works. The current systems rely on data such as crime statistics and local demographics. The problem with these statistics is that they are difficult and expensive to gather and not regularly updated.

By contrast, mobile phone operators can collect data about the owners such as their gender, age and so on and then monitor the location of the phones in real time. It’s not hard to imagine that this kind of data might significantly improve the accuracy of crime prediction models.

That’s exactly what Bogomolov and co set out to show. These guys used a dataset about mobile phone users in the centre of London, which they obtained from Telefonica, a European mobile phone company which owns the O2 service in the UK.

This dataset contains all kinds of information about the users: their age, gender, home location and so on. But it also contains their location at various instants in time. To find out whether this could improve crime prediction, Bogomolov and co used the crime statistics and demographics from one period of time to train an algorithm to predict crime rates in the next period of time.

The results were pretty good. In that case, the algorithm was able to predict whether a given area would be a crime hotspot or not in the next month with an accuracy of about 62 per cent.

Then they added the phone data into the mix and used the resulting dataset to retrain the algorithm. That significantly improved the accuracy to 68 per cent. That’s an increase in accuracy of 6 per cent — not revolutionary but still significant.

Their analysis shows that some mobile phone data is more important than others. For example, the data relating to whether or not the phone owner was at home, was particularly strongly correlated with crime patterns.

That’s an interesting result that should help law enforcement authorities use their resources more effectively. “The proposed model could be used to predict new crime occurrence areas that are of similar nature to other well known occurrence areas,” say the team.

It should also help limit the cost of these kinds of predictive systems because conventional demographic data is expensive to gather. By contrast, mobile phone data is extremely cheap.

But there are also issues to guard against. Perhaps the most controversial is privacy. Telefonica says the data sets are entirely anonymised so that it shouldn’t be possible to identify any individual from the data.

If you don’t find that re-assuring, you won’t be alone. There are numerous examples of seemingly anonymised data being used to reveal people’s identities.

And while this approach is place-centric, focusing law enforcement on locations rather than individuals, it’s not hard to imagine authorities taking this kind of approach further by applying crime prediction techniques to the behaviour of individuals.

Perhaps Minority Report wasn’t so far-fetched after all.

Ref: arxiv.org/abs/1409.2983 : Once Upon a Crime: Towards Crime Prediction from Demographics and Mobile Data