University of Pennsylvania criminology professor Richard Berk tackled Philadelphia’s probation-and-parole challenge with a computer-modeling technique called “random forests.” I had the pleasure of being a part of the Penn Criminology masters program when Professor Berk joined the Penn Faculty. I also worked with Dr. Berk while he did some forecasting for the Pennsylvania Board of Probation and Parole.
Here’s how random forest works, according to the Pennsylvania Gazette:
Berk gathered a massive amount of data about 30,000 probationers and parolees who’d been free for at least two years. He fed it into an algorithm that randomly selected different combinations of variables, and fit the information to a known outcome: whether someone had been charged with homicide or attempted homicide in that time frame. The algorithm repeated this process hundreds of times, producing a “forest” of individual regression trees that took arbitrary paths through the data. For instance, one tree might begin by considering parolees’ ages, then the number of years that had passed since their last serious offense, then their current residential ZIP code, then their age at the time of their first juvenile offense, then ZIP code (again), then the total number of days they had been incarcerated, and so on, creating a sort of flow chart that sorts any given individual into a category: homicide, or no homicide. Another tree would follow the same procedure, but using different combinations of variables in a different order.
To test the predictive power of this forest, Berk then fed it data on 30,000 different cases—whose outcomes were also known, but which had not been used to build the model. Each was assessed by every tree in the forest, which cast a “vote” on the likelihood that the individual would try to kill again. Those votes were tabulated to generate a final forecast for each case. Importantly, the forest is a black box; there’s no way to know how—let alone why—it arrives at any given prediction.
Assessing a prediction’s value is tricky. Out of the 30,000 individuals in the test sample, 322 had actually been charged with homicide or attempted homicide within two years. So simply predicting that any given person would not kill again would make you right 99 percent of the time. But that would prevent no deaths. A standard logistic regression using the same data, by comparison, fingered two out of 30,000 subjects as likely to commit murder, and it was right about one of them. Not very impressive, but at least it might have saved one life.
Berk’s algorithm was in a different universe. It forecasted that 27,914 individuals would not attempt murder within two years, and it was right about 99.3 percent of them. It identified 1,764 as at risk for killing, 137 of whom in fact faced homicide charges. Generating a prediction for any given individual, using data already available to criminal-justice decision-makers, took “just 10 or 15 seconds,” according to a subsequent review.To read more CLICK HERE