Joins from Lyst to apply deep learning to the huge issue of fraud detection and false declines.
I've been following Ravelin's progress with great interest over the last two years, and have been very impressed with the maturity and intelligence of their approach. I'm thrilled to join as Head of Machine Learning, and I'm looking forward to working closely with the team to further advance their detection strategies and take the fight to fraudsters!
Ravelin uses a variety of techniques, including Machine Learning (ML), to detect fraud. Until now, online businesses have had to manually define and balance different "Rules" which decide which customers or orders are legitimate or fraud. With Machine Learning we can use live and historical data to produce "Models" which do this all automatically with a much greater capacity for detecting patterns and changes. This ends up being not only faster and more accurate, but much more maintainable as 10 rules are easy to juggle but 50 or 100 or 1000 very quickly get out of hand and impossible to truly understand.
There is also the problem of attrition; over time rules become less relevant as the environment changes and new rules supersede their predecessors. Without careful custodianship rules quickly become an unmaintainable tangle. The same is not true of ML models. If the assumptions or data change, then we simply retrain to reflect the change in environment.
My particular specialisation is deep learning. Deep learning has gained massive traction over the last five years due to its capability to learn complex patterns and because it helps practitioners construct bespoke models which are ideally suited to the problems they are trying to solve. Deep learning has many potential applications in fraud from learning customer similarity to building fraud networks and distilling complex structures down to useful representations. Let us consider an example of textual similarity.
Traditional models treat text as an unordered set of words. Decisions are made based on the existence (or non-existence) of words. We can think of this approach as symbolic because there is no context or understanding of a word besides the fact that it exists. Deep approaches typically learn contextual semantic representations for words. With such representations, we can easily determine that the terms cat and cats or cat and dog are related. This would be much harder in a symbolic model.
These deep textual representations could allow us to catch fraudsters more easily by matching variations in the names and email addresses with which they register. This takes us beyond the current best practice whereby the attributes of an email address, which may indicate fraud, are manually curated. It also takes us closer to mimicking a part of the human ‘intuition’ in fraud assessment but does so at much greater scale and speed.
This is just one example of where deep learning can make a transformative difference in not just detecting bad transactions but also on the converse problem of allowing more good transactions.