Can future knife crime perpetrators be predicted using administrative information already known to the police with machine learning? Reported crime data, recorded between 29th April 2014 and 14th June 2019 within the Thames Valley region, was used to build a model to test this theory. 

Feature variables about the potential perpetrator were created from this data set. Using social network methodology, feature variables were also established for potential co-offenders of the perpetrator. 

A random forest model was generated to predict if a perpetrator in any reported crime would commit a knife crime within one year. 

This predictive model follows a number of studies in the UK, and abroad, examining the social network indicators of violence using police data. These earlier studies have explored evidence of individuals with previous criminal histories, associates within their social circles who have violent or criminal backgrounds, and a history of possessing a weapon, as indicators of whether an individual is likely to engage in knife crime. 

Levels of knife crime have been described as at ‘epidemic levels’ in the UK, and, therefore, there is an urgent need across our society to explore the factors underpinning this insidious form of criminality. It is vital that policing looks at methods which may be deployed to both predict and prevent individuals getting sucked into becoming perpetrators of knife crime or, by extension, becoming victims of it themselves.  

Please read a full account of this study within the TVP journal (Pages 10-22).