The use of credit cards and other cashless or digital payment methods has become the norm for consumers all over the globe, and the strong surge of online buying during the pandemic has further boosted this decade-long trend. However, behind the convenience of ’click and pay’ there are also risks, such as fraud and related losses, which are mostly borne by the card companies. How to combat credit card fraud, or at least limit the damage? Researchers at USI have recently completed a project, performed in collaboration with a major credit card company in Switzerland, that has developed an innovative probabilistic model for the efficient detection of fraudulent transactions.
It is estimated that worldwide, losses to card companies amount to more than 7,15 cents for every $100 of credit card transactions - a figure that translates to about $25 billion in real terms, and it is estimated that this figure could double by 2025. We had reported on this in 2016, when Dr Bruno Buonaguidi , a researcher at USI under the supervision of Professor Antonietta Mira , won a grant from the AXA Research Fund to lead the project (see Quicklink in the sidebar). The project has now been completed and the results published in the scientific journal Bayesian Analysis. But what does the solution devised at USI consist of?
"We have developed a system that works in two phases, the training phase followed by the classification phase," explains Dr Buonaguidi. "Our algorithm first studies the behaviour of credit card users, to understand their purchasing habits, and translates this into certain statistics, such as average time spent, average amount, geographical distribution of purchases. At the same time, it also studies the characteristics of transactions that we know to be fraudulent (because they have been reported by credit card holders, for example). The algorithm returns, for each credit card user, a personalised threshold that will be used to declare the nature of the transactions. In the second step, the algorithm calculates for each new transaction of a credit card user the probability that this transaction is fraudulent. When this probability exceeds the threshold referred to in the previous point, the fraud alarm is raised - for example, if the time between two transactions is very close together, with high amounts and in a place other than where purchases usually take place, the model will return a high probability of fraud that is likely to exceed the threshold".
According to Professor Antonietta Mira , who heads Data Science Lab at USI and who supervised the research project, "the prospects for application by our project partner, as well as other credit card companies, are good. In fact, by applying our algorithm to real data and comparing it with other machine learning methodologies currently used in the industry, we found that the percentage of false positives (i.e. legitimate transactions classified as fraudulent) is on average similar, while the percentage of true positives (i.e. fraudulent transactions classified as such) is on average higher in our model."
On the subject of algorithms the television programme of RSI ’Tempi moderni’ aired a report, featuring the comments of Dr. Buonaguidi and Prof. Mira, and mentioning also the USI project for the ’smart’ deployment of defibrillators in Ticino (see embedded video below).