06-07, 10:20–11:05 (Europe/London), Grand Hall
This talk will present Monzo's exploration of multi-task deep learning to enhance our real-time fraud detection systems. I will outline the challenges of card fraud detection, and explain the limitations of traditional gradient boosted decision tree models in terms of generalisation to rare fraud subtypes. This will motivate the use of multi-task learning, which leverages shared dense representations across fraud sub-tasks. By consolidating multiple specialist learners into a single model, we observe improved performance on less prevalent fraud types, leading to better generalisability, scalability, and robustness. I will also share results from testing multi-task models within our fraud detection infrastructure.
Fraud detection is a complex problem due to the constant evolution of fraudulent behaviour, significant data imbalance, and the requirement for real-time decision-making. Accurate detection of fraud and financial crime is crucial for protecting customers and maintaining trust in the banking system. Traditional fraud detection often relies on binary classification models using tree-based algorithms. While these models offer good predictive performance and scalability, they can struggle to capture shared information across different types of fraud. This often results in the need for multiple specialist models, each requiring individual maintenance and retraining.
Multi-task learning, a deep learning approach, offers a potential solution by exploiting the commonalities between related fraud problems to improve overall prediction accuracy. Multi-task learning is particularly relevant where multiple prediction targets share underlying patterns. In fraud, different sub-types (e.g., identity theft, account takeover, coercion) frequently exhibit overlapping characteristics. A model trained on multiple signals simultaneously may be better at identifying subtle patterns that individual models might miss. Our hypothesis is that this should lead to increased generalisation, allowing multi-task models to adapt more effectively to new fraud patterns and reduce maintenance overhead.
In this talk, I will detail how we have tested this hypothesis at Monzo by applying multi-task learning to the problem of unauthorized card fraud. I will discuss the models we developed and the results we have observed in controlled offline settings..
No previous knowledge expected
