Ravelin's head of machine learning in conversation
Getting past the hype
There is often more hype than information available about machine learning. That’s why we were joined by Eddie Bell, the Head of Machine Learning at Ravelin to discuss the use of machine learning in fraud prevention. Eddie defines machine learning simply as a set of methods and techniques that let computers recognise patterns and data, and make predictions - read more about this here.
As machine learning models are normally built on historical data, it can be complex to build a model for a client without a transaction history. Eddie tells it is always better to build a customer-specific model but sometimes it might not be possible. Ravelin uses two approaches to tackle the problem. Either a generic model built on several clients and countries can be used which often delivers good performance or instead of one big model, micro models can be applied. A bigger model would be split into smaller components which could be detecting email, location or transaction history. The advantage of micro models is that fraudulent patterns often look similar and a feature can be trained generically.
Why we choose the approaches we do
With machine learning models, customers often want explainability. The simpler the model, the easier it is to explain - but the worse it is at predicting fraud. The machine learning team at Ravelin wants the best performance for their clients and explainability in terms of being fully aware of what is happening. The approach used by Eddie’s team is to change the input used in the model and see how it changes the output. Through this process they are able to probe what is happening. If the model is favouring email addresses, this can be presented to the client to explain why certain decisions are made. Another technique to add explainability is unsupervised learning.
The team will not look at labels or chargebacks but only the customer; is there anything strange about their data? Maybe they have added 27 credit cards or had ten transactions in a day. They might not be fraudulent but there is something suspicious about their transaction or card behaviour that can be flagged up to the client.
It seems like using machine learning is a seamless way of preventing online fraud. Is there anything machine learning is not good at? Eddie describers catching fraudsters as cat-and-mouse game - his team is constantly on the lookout for new signals indicating fraud. There will always be a tiny percentage of fraud getting through but this is not something humans would be able to detect either. If it’s something that a human could do, the same can be done with machine learning but faster and more efficiently.