A large internet payments company.
The company specialises in mobile payments, processing more than eight million transactions every month with a collective value of over £1 billion a year. Although fewer than 0.0001% of transactions were fraudulent, the costs of fraud and of the fraud detection team were among the company’s largest expenditures.
The company had a rule-based fraud detection system running onsite, operated by a team of people. The team struggled to keep up with the growing volume of transactions, the huge amount of data and the evolving tactics of fraudsters. The rules applied became more and more complex and staff costs were growing, yet fraud detection had plateaued. The company asked Faculty to develop an artificial intelligence (AI)-based approach to help it detect fraudulent transactions.
AI is well suited to processing huge amounts of data and to be able to evolve automatically to handle the changing nature of fraud.
Firstly, we worked with the existing team to understand the domain and identify known types of fraud. Secondly, we removed noise from the data, and applied cutting-edge techniques (t-Distributed Stochastic Neighbour Embedding, t-SNE) to overcome the technical challenge of training a model on a data set in which only a tiny percentage of data points are anomalies that need to be captured.
2D Projection of the transactional data by t-SNE. The scatter plot and KDE plot illustrate how the algorithm has successfully distinguished between the frauds on the periphery of the circle and the normal transactions towards the centre.
The greater accuracy of the AI-based system meant that 93% of fraudulent transactions that would have slipped through the previous system can now be detected. Fewer legitimate transactions are blocked by the new model. Revenues have increased and customer satisfaction has improved. Once productionised, this model will allow the organisation to drive operational efficiencies that will save approximately a fifth of the overall salary budget.