Computer algorithm accurately identifies people felons

Recently in the journal Science Advances published the results of research that has the potential to transform criminal justice. A team of researchers from Stanford University and the University of California, Berkeley, found that computer algorithms rather people determine who of the accused may commit a crime again.

In the new work, scientists say that even untrained people can fairly accurately predict the risk of re-offence by the person: they only need to know a few simple variables. But the criminal justice system who are not eligible for error, works with a large number of parameters: in this case, even the professional to draw conclusions much more difficult. Recently, it became clear that the algorithms deal with the prediction of a possible relapse is much better than people.

“Proven risk assessment tools [re-offending] can help professionals in the field of justice to make better informed decisions, – said Charades Goal (Sharad Goel) from Stanford University. For example, these tools can help judges to identify and release people who pose little risk to public safety.

But, like any other measures, risk assessment tools need to work under human supervision and be combined with sensible policy to criminal justice reform was fair and effective.”

Explain that the assessment tools, based on algorithms widely used in the United States in areas such as health care, banking and even for admission to the University. They have found their application in criminal proceedings, where they help officials to analyze data in making decisions.

But in 2018, researchers from Dartmouth University have questioned the accuracy of this method and its relevance for criminal justice. During the research they studied a thousand cards of defendants in criminal cases, containing a brief information about these people.

Each of them included five factors on which you can draw some conclusions about the person: gender, age, current criminal charges, as well as the number of previously committed crimes in adult and juvenile age. Information proposed to evaluate 400 volunteers and asked to guess whether the accused committed another crime within the next two years. The organizers had on hand information about which of the accused committed the crime and who is not. Therefore, every decision of the participants of the experiment they were accompanied by a comment relating to the correctness or incorrectness of their answers.

the Likelihood whether the accused a repeat offender, was assessed and the prediction algorithm called COMPAS. The basis for them was taken by the same five parameters listed above.

as a result, the people and the forecasting system gave accurate answers around two thirds of the cases: and if the volunteers were right in 62% of cases, the algorithm is 65%.

These results, according to Dartmouth researchers have questioned the value of risk assessment tools and algorithmic forecasting.

the Study caused a wide resonance in the press. Many are talking about that makes no sense to use algorithms, whose accuracy is comparable with a human.

the Authors of the new study decided to expand the original work. And in addition to the five evaluation factors, which are usually used COMPAS, they introduced another 10, giving the algorithm and people appreciate them. In particular, the evaluation system was to take into account such variables as the employment status of the alleged offender, the use of psychoactive substances and his mental health.

the Methodology of work was also revised slightly: for reasons of clarity, the organizers have not always confirmed the correctness of the judgments of volunteers. Thus, people assessed the risk of recurrence of the crime, not based on their previous responses. Recall that judges and other judicial decisions in similar circumstances: no one ever gives them immediate assessment of the correctness of their judgments.

On the result of the estimation algorithm outperformed the accuracy of people.

So, the analysis of the COMPAS dataset and the same method augmented with ten drivers, has reached 89% accuracy. Volunteers also showed good results when they received feedback about the correctness of their answers. The accuracy of their estimates was reached 83%. But without such “help” the correctness of the answers of the volunteers was only 60%.

As the scientists explain, the results show that people can predict the recurrence on a par with the statistical models if the data for the analysis contain some simple factors. In this case, the accuracy of computer algorithms and people who did not receive feedback, are largely similar. But when working with large amounts of data, the algorithm is more accurate than the human forecasts. As the researchers note, the fact that the models better take into account more information than people.

So, if the advanced risk assessment tools will continue to improve, it can lead to that critical decisions will be made more accurately. This is important, because experts in the field of justice every day decide what kind of person can be rehabilitated in society and not in prison? One of them can be enclosed in a prison with low security, and any institutions with a strict regime? Which prisoners can be released on parole, and what to leave prison before the end of the sentence?

we will Add that earlier the authors “News.Science” (nauka.vesti.ru) wrote the application that will find the sniper by the sound of the shot, and also that the offender can give eye movements.

Text: To.Science