Neural net ‘precrime’ system suitable for prospective employees, say Russian scientists
Tue 11 Oct 2016
Researchers at Tomsk University have created a computer program that can be used to predict different aspects of criminal behavior using Artificial Neural Network (ANN) technology, and have suggested that the technique could be adapted as an effective tool in selecting suitable employment candidates.
The team was able to divide survey respondents into classifications based on the ANN program’s analysis of dominant personality characteristics, which were correlated to the type of crime they had been jailed for in 80% of cases.
The objective of the study [Russian] was to ascertain whether or not prisoners could be classified based on the survey results using a specially-developed ANN program, which would separate respondents based on dominant characteristics; and to see if those dominant characteristics were related to the type of crime for which the subject had been convicted.
The research team developed a survey that was distributed to 180 prisoners jailed for offenses including murder, grievous bodily harm, theft, and drug-related crimes. Completed surveys were then analyzed using two different artificial neural network (ANN) programs, to measure personality characteristics in an attempt to classify prisoners based on their predilection for a certain type of crime.The study found that in 80% of cases, the scientists could accurately predict what crime a prisoner had been convicted of, based on the dual ANN program analysis of the personality survey results.
The survey attempted to cover a large set of personality characteristics, while the Artificial Neural Network divided responses into two separate, scaled two-factor classifications. The first measured respondents on a scale of domination to subordination and the second, friendliness to hostility.
Project lead Michael Golovchiner, Associate Professor in the department of Applied Mathematics at Tomsk University, said, “We are interested in the problem of predicting behavior, which is primarily determined by the individual’s temperament and character.” He continued, “The question of whether there is a set of characteristics that distinguish criminals convicted of various crimes is a task well-suited to ANN, because this is an issue of classification.”
“The training and testing revealed two characteristic factors: ‘domination-submission,’ and ‘friendliness-hostility,’ and the error rate of classification was about 20 percent, indicating a relatively high level of accuracy.”
Golovchiner also noted that the survey/analysis model could be used in a variety of settings, using the ANN program to predict other types of behavior (in addition to criminal actions) based on a number of different characteristics. For example, an employer could use a combined survey/artificial neural network analysis to create a deeper understanding of candidates for key positions at a company.
