UK police use computer algorithm EBIT to increase the efficiency of crime investigations by about a factor of 1

Kent police in southeast England, are experimenting with computer algorithms to improve the efficiency of case processing. Such algorithms can assist in determining which cases are more effectively solvable. Police forces using this algorithm have roughly doubled the efficiency of their investigations, saving significant time and money.

Some are concerned that the algorithm may interfere with subjective human judgment and cause some cases that could be solved to be overlooked. Therefore, the tool is currently only used to assess areas such as assault crimes and public order crimes but may be extended to other crimes in the future.

When the police receive a report of a crime, they will usually send someone to the scene to verify the situation. When a genuine crime is encountered, an immediate arrest is made. However, not all reported crimes are worthy of further investigation. In many cases, the police can only use their own experience to decide whether further investigation is necessary.

But thanks to EBIT, it is clear that the police here can handle more cases now. The Evidence-Based Investigation Tool (EBIT) algorithm, which yields a probability score for the solvability of a crime.

Kent Police was the first police force in the UK to experiment with predictive policing. Before using EBIT, they had also experimented with another algorithm to predict crime. This proprietary machine learning algorithm, provided by the American company PredPol, is used to predict potential crime hotspots over five years, suggesting areas where crime is likely to occur. Kent Police previously spent £150,000 a year on a contract with PredPol, but because the technology required police to visit 520 hotspot boxes a day, the police were only able to visit an average of 86. It, therefore, lacked comprehensive data to back it up, so last year Kent Police stopped using the technology.

After practice, Kent Police decided that EBIT was more operational than the algorithm provided by PredPol.

Assault crimes and public order offenses account for around a third of all crimes in Kent, such as threatening someone on the street. And having used EBIT to assess the solvability of assault crimes and public order offenses, it can reduce the 75% arrest rate, which would otherwise rely on human judgment, to 40%, all other data remaining the same. This will allow police forces to prioritize more valuable investigations; after all, cases are never-ending and impossible to solve in their entirety.

McFazion, of Cambridge University, trained the EBIT algorithm with data from thousands of assault crime and public order cases. Among the eight factors that determine whether a case is solvable is the availability of witnesses, the presence of surveillance footage or the existence of specific suspects, and so on.

Because these factors change over time, EBIT recommends sampling one or two low-solvability crimes from the algorithm's results each day. This is a blind test of the algorithm's effectiveness.

Because the technology's predictions are based on past survey data, the algorithm can reinforce any bias contained in the decision. For example, EBIT found that some areas had no surveillance video, resulting in a low rate of police investigations there, a factor that could lead to a possible security risk for the population here.

"When we train the algorithm on data about historical arrests or crime reports, any biases in the data will go into the algorithm, and the algorithm will learn those biases and then become more effective," says Loftus of Stanford University in California.

So blind testing is an effective way to help solve the problem of bias, but there will always be some problems that are not easy to solve, such as those in communities where there is distrust of the police and a large number of unreported crimes, so there is a lack of valid data to support the problem.

Kent police chief John Phillips said that although there is a certain lack of data on the sample of cases, the type of crime used by EBIT, does not affect the validity and reference value of the tool.