We recently caught up with George Mohler, Chief Scientist at PredPol, Inc and Assistant Professor of Mathematics and Computer Science at Santa Clara University. We were keen to learn more about his background, the theory and technology behind predictive policing and the impact PredPol is achieving ...
Hi George, firstly thank you for the interview. Let's start with your background and how you became interested in predicting crime hotspots...
Q - What is your 30 second bio?
A - I am Chief Scientist at PredPol, Inc and Assistant Professor of Mathematics and Computer Science at Santa Clara University. Prior to joining the faculty at Santa Clara University I was CAM Assistant Adjunct Professor of Mathematics at UCLA from 2008 to 2010. I received a B.S. in Mathematics from Indiana University and my Ph.D. in Mathematics from the University of California Santa Barbara.
Q - How did you get interested in Data Science and Machine Learning?
A - I became interested in data science rather late, during my postdoc at UCLA. Prior to that I was working on computational methods for variational models of polymers in graduate school. When I joined the crime modeling group at UCLA, I started to work on similar types of optimization problems, but applied to spatio-temporal crime patterns. We had a large dataset provided by the Los Angeles Police Department and I was interested in understanding the statistics of crime hotspots and how they could be predicted.
Q - What excites you most about bringing Data Science and Policing together?
A - Some of the models we use at PredPol are self-exciting point processes that were originally developed for modeling earthquake aftershock distributions [Marsan and Lenglin, 2008]. The fact that these point process models fit earthquake and crime event data quite well is, by itself, a cool result. However, in the context of policing we can actually send police into the hotspots that we predict in order to prevent crime. So not only does predictive policing present an interesting modeling problem, but the models then have a societal impact that can reduce the risk that one's car is broken into or that they are a victim of gun violence.
Editor Note - If you are interested in more details on the research underlying the models, the original academic paper is very insightful. Here are a few highlights:
- Criminological research has shown that crime can spread through local environments via a contagion like process. For example, burglars will repeatedly attack clusters of nearby targets because local vulnerabilities are well known to the offenders
- Self-excitation is also found in gang violence data as a gang shooting may incite waves of retaliatory violence in the local set space (territory) of the rival gang. The local, contagious spread of crime leads to the formation of crime clusters in space and time. Similarly the occurrence of an earthquake is well known to increase the likelihood of another earthquake nearby in space and time
- Mohler and his fellow authors propose (and demonstrate!) that self-exciting point processes can be adapted to capture the spatial-temporal clustering patterns observed in crime data. More specifically, spatial heterogeneity in crime rates can be treated using background intensity estimation and the self-exciting effects detected in crime data can be modeled with a variety of kernels developed for seismological applications or using nonparametric methods
Editor Note - Back to the interview!...
Very interesting - fascinating that the earthquake models can be applied to crime event data! Let's talk more about the technology that you have built at PredPol...
Q - Firstly, what specific problem does it solve?
A - Police departments have limited resources. Officers have large beats to cover and limited time when they are not on a call to service. So during a given shift in a given beat, an officer might only be able to patrol k hotspots. From a prediction perspective, we would like to flag for patrol the k hotspots that are most likely to have crime in the absence of police. This means that hotspot policing is actually a learning to rank problem.
Q - How does the technology work?
A - PredPol is a SaaS company and officers access the software on a computer or smart phone with an internet connection during their shift. They then pull up a UI that includes a map with the hotspots (150m x 150m) displayed. The idea is that the officers then make extra patrols in those areas when they are not on a call to service. The key with this sort of technology is to make it as simple and easy to use as possible, because police have a very difficult, dangerous job and they don't have time to mess around with complicated software in the field.
Q - That makes sense... And what is your favorite example of how PredPol is having real-world impact?
A - We have run randomized controlled trials to measure accuracy of the PredPol algorithms and impact on crime rates. These are necessary, because without them it is impossible to determine whether a crime rate increase/decrease is due to the technology and its use or because of some exogenous factor. But my favorite examples are at the scale of individual hotspots. For example one agency had a guy stealing cars with a tow truck. So the police put a decoy car with a GPS tracker in one of the PredPol hotspots and sure enough he came and towed it away (and the police were able to catch him).
That's a great example :) It sounds like you've already had significant impact with PredPol - let's talk a little about the future...
Q - What research areas would you like to explore more going forward?
A - We still do a lot of work on improving our algorithms at PredPol, both in terms of bringing in new modeling approaches and also exploring what loss functions make sense for policing (and how to optimize them). However, predictive policing is not just about designing accurate algorithms. Ultimately the software has to be used by police in the field and so human-computer interaction is really important. We are exploring ways in which the software can increase officer patrol time in hotspots while still fitting seamlessly within their existing practices.
Q - And finally, any words of wisdom for Data Science / Machine Learning students or practitioners starting out?
A - Many universities have courses in machine learning and some are starting to have degrees specifically in data science. But there are many ways to learn data science on your own. I think Kaggle is a great way to start out and there are some entry level competitions that walk you through some of the basics of data science. Coursera has free courses in data science and machine learning; I took Bill Howe's "Introduction to data science" class over the summer and thought it was really well put together. I recommend to my students that they try to do an internship or REU in data science if they are interested in pursuing a career in the area.
Readers, thanks for joining us!
P.S.If you enjoyed this interview and want to learn more about
- what it takes to become a data scientist
- what skills do I need
- what type of work is currently being done in the field