Stephen Whitworth is one of the co-founders here at Ravelin, where he works as a Senior Data Scientist. He previously worked at Hailo before making the move to become a co-founder at Ravelin with four other colleagues. Stephen was recently nominated as a Forbes 30 Under 30 and recognised for his impressive work in machine learning.
This interview is part of a new series that we’re introducing to get to know the team and their roles at Ravelin.
Can you introduce yourself and tell us a little bit about your role?
I’m one of the co-founders at Ravelin. I work as a machine learning engineer within our detection team, where I’m responsible for insuring that the fraud detection that we give back to clients is always improving. This is more than just looking at machine learning; it’s creating strategies around rules and graph networks and how they all intertwine to give the best detection possible.
On a day-to-day basis this means building things, so to say! Building infrastructure to pick out signals and features from customers. Trying to identify good customers. There’s quite a bit of infrastructural work that we do on extracting the data and getting it into the right format, where we then train the models and make sure that they are performing as well as they can be.
We focus a lot of our time on evaluation, to ensure it’s easy for our team of data scientists to understand how good these models are working and ensure that it’s as easy as possible to deploy them into the real world.
Why do we use machine learning at Ravelin?
There’s a multitude of reasons why Ravelin use machine learning. Our business models are essentially built to spot and stop fraud on behalf of clients. One of the reasons machine learning provides great benefits in fraud detection is that it can quickly learn from examples of what fraud looks like in the past, and use that to pick apart different types of fraud that has been seen previously.
Ravelin’s clients have large datasets, and fraud looks very different for each one. Without machine learning, we would have to have fraud analysts constantly trying to spot trends in each client’s data.
Human intervention and manual review may mean overreacting and blocking some orders which may not always be the right decision to take. Machine learning can understand the weighings of different types of fraud and deal with them accordingly. It learns things that are harder, or may be counter-intuitive for humans to spot.
Can you tell me a little bit about scores? What contributes to one and how is it evaluated?
The machine learning score is essentially a number between 0-100. This score portrays our estimate of how fraudulent a customer is. To score a customer, we extract hundreds of signals about a particular customer: the number of cards they have added recently, the distance between the letters on the keyboard in their email, whether they have copied and pasted a card, etc.
After we’ve extracted all of these features, we run them through a model. You can think of the machine learning models that Ravelin uses as a series of questions that are asked of the features. For example, does the customer have an account less than two weeks old? Have they ordered in a historically fraudulent area? Repeatedly asking these questions will give you a very specific understanding of how fraudulent you think that customer is based on all the historical examples we have seen in the past. Based on the estimates we’ve seen, we output a number between 0 and 100, 0 being the most likely to be genuine, and 100 being the most likely to be fraudulent.
Are Ravelin’s rules and model customisable per client?
Models are the core of our solution. We train new models for each client on a regular basis and deploy new models when performance has improved - a key factor in this is how often the client submits chargebacks to us. We are happy to build bespoke signals for a clients business when appropriate. Manually reviewing customers in the Ravelin dashboard is also available to impact the model performance. Outside of the model, there is a rule system you can use to enforce business policies and the boundaries of acceptable behaviour.
What are some characteristics of fraudsters that you’ve come across?
I would say aggression of payment and behavior. People who are making purchases at a much higher rate than you would usually accept are usually the ones to watch. And adding lots of cards, ordering repeatedly, making expensive orders, ordering items that have higher resale values and so on. Generally being bolder than usual.
And finally, what’s your favourite thing about fraud detection?
Using machine learning to stop crime and stop innocent people losing money. This makes me proud of what we do at Ravelin. There are a lot of ethical implications to it, and catching out the bad people and protecting the good is something we value.