How Ravelin’s graph database transforms fraud detection for our clients
At Ravelin we’ve been primarily known to date for the application of machine learning and AI techniques to uncover fraudulent patterns in large datasets. This is first amongst a number of techniques that we use to find fraudsters fast and stop the transaction before payment. However we have been putting increasing effort into graph network techniques to uncover fraudulent and/or bogus accounts.
We’ve been working on this under the hood a little bit, working directly with clients and surfacing networks of fraud in their customer bases. We are shortly going to roll this out to all of our clients so we wanted to explain how it works and some of the approaches that we have taken to do this — approaches that we think are ground-breaking both in the speed and efficiency with which they can uncover and continue to uncover fraud networks.
Blazingly fast fraud network detection
Ravelin visualises these graph databases directly in the analyst dashboard. Users can immediately see suspicious nodes with multiple cards associated with a single user. Traditional rule-based fraud systems are quite good at catching these single instances of users. But what they cannot do is find connections that may already exist in a network.
To find these connections we needed a graph database. Our aim was to find a tool that would allow us to quickly and easily map out clients’ data and expose connections based around the deliberately limited set of parameters that we set.
There were a number of challenges in doing this, most significant of which was the system load. When we started to investigate graph databases we found ourselves disappointed; there aren’t many production-ready graph databases out there. The ones we tried were difficult to use and slow to load data. We tried them out, but as the hours ticked by waiting for our test data to load we began to look for another solution.
Graph databases for non-specialists
We ended up building our own proprietary graph database from the ground up, optimising for the specific types of questions we needed answers for. The result is that we can generate a full graph for any given customer in single digit microseconds.
Using these techniques at this speed is transformative in fraud detection. Uncovering connections and networks is now cheap from a resource point of view and instant from a client point of view. Graphs can be generated on the fly, with minimal resource overhead — moving the generation of the graphs out of the hands of data specialists and into the hands of fraud analysts.
This is huge step change. At client sites where we have trialled this, we have uncovered — and our clients have eliminated — thousands of fraudulent accounts instantly. Graph databases allow us to block suspect and bogus accounts before they have taken any fraudulent action. While this has been possible before, being able to create and query these networks on-the-fly by non-data specialists completely changes the game in how graph databases can be used in this field.
This is a tremendous ride-along with our core machine learning fraud detection which is reliant on actions taken by fraudsters for us to score and ultimately stop them; graphs allow us and our clients to find multiple bogus actors for every single one we prevent through scoring.
Any client implementing the graph database approach within Ravelin should see a dramatic clean-up of active and dormant bad actors in their own network. We are also able to track this activity across clients. Like all features in Ravelin we have put a huge emphasis on its discoverability and usability for the analysts. Consequently, perhaps most exciting part of these graph databases is that analysts can apply this technique in ways we have not imagined yet — that’s the beauty of democratising the access to a powerful technology. We can’t wait to see what develops.
Graph Network Analysis is available today at no extra cost currently as part of the core Ravelin service.