AI Case Study
Durham Police Constabulary improves custody decisions by predicting whether offenders will re-offend using machine learning that results in 98% avoidance of false negatives
Durham Police Constabulary predicts whether an offender is at high-risk of offending. If so they are charged and those with a lower risk are offered a rehabilitation program and forgo a criminal conviction. The modelling used random forest machine learning based on comprehensive profiles of individuals from Experian such as geography, family composition, occupation and even names linked to ethnicity. This resulted in prediction accuracy of over 60% and 98% accurate in avoiding false negatives.
Public And Social Sector
"Durham police trained a machine learning system called HART ("Harm Assessment Risk Tool") to determine those who are at risk of reoffending. Working with Cambridge University and overlaying Experian data on the UK population, the model determines which offenders to spend more time to reduce the chance of reoccurrence.
The aim of HART is to categorise whether in the next two years an offender is high risk (highly likely to commit a new serious offence such as murder, aggravated violence, sexual crimes or robbery); moderate risk (likely to commit a non-serious offence); or low risk (unlikely to commit any offence)."
To better optimise limited police resources, the Durham police wanted to know to know whether someone should be charged with an offence based on whether they are likely to re-offend. "'High-risk' offenders are charged. Those with a 'moderate' risk of re-offending are offered the option of joining a rehabilitation program; if they complete it successfully, they do not receive a criminal conviction."
"An independent validation study of HART found an overall accuracy of around 63%." The model was 98% accurate at avoiding false negatives.
"Called the Harm Assessment Risk Tool (HART), the AI-based technology uses 104,000 histories of people previously arrested and processed in Durham custody suites over the course of five years, with a two-year follow-up for each custody decision. Using a method called 'random forests', the model looks at vast numbers of combinations of ‘predictor values’, the majority of which focus on the suspect’s offending history, as well as age, gender and geographical area. T
The "technology uses 104,000 histories of people previously arrested and processed in Durham custody suites over the course of five years, with a two-year follow-up for each custody decision.
Durham Police is feeding Experian's 'Mosaic' data, which profiles all 50 million adults in the UK to classify UK postcodes, households and even individuals into stereotypes, into its AI 'Harm Assessment Risk Tool' (HART). The 66 'Mosaic' categories include 'Disconnected Youth', 'Asian Heritage' and 'Dependent Greys'.
As well as using basic zipcodes, a wide range of sensitive "predictor values" are gathered, aggregated and analyzed, such as:
Family composition, including children,
Family/personal names linked to ethnicity,
Online data, including data scraped from the pregnancy advice website 'Emma's Diary', and Rightmove [UK real estate site],
Child [support] benefits, tax credits, and income support,
[Children's exam] results,
Ratio of gardens to buildings,
Gas and electricity consumption."