We are the hosts of a regular artificial intelligence meetup called Applied AI, that showcases the latest and greatest in AI tech. If you'd like to learn more or join, please visit our meetup page.
On 19th January, Ravelin sponsored the second Applied AI Meetup at the Google Campus in London. The theme around these meetups is product applications of AI in commercial, not research or academic, spaces (you can sign up here). I spoke about how to build products where the data and intelligence you can gain from them forms the core of the product, and the heuristics you should keep in mind when building them.
At the beginning of the talk, I spoke about the causes of the problem that Ravelin is trying to solve. Card not present fraud is growing very quickly with the move to mobile. The huge uptake of ‘on-demand’ services, powered by mobile devices, has meant that there simply isn’t the time to manually review lots of orders for fraud anymore - you need to make a decision instantly. The plethora of user data floating around the dark web due to data leaks from web applications means that the barrier to entry to using stolen credit cards is lower than ever.
Consumers aren’t often aware of the big impact fraud has on merchants. Fraudsters are rare in the population, but they have an extremely disproportionate negative impact on a business’ bottom line. The above image illustrates it well. See if you can find the single N in 999 M’s. Finding a needle in a haystack, or the one bad fraudster in a sea of good people, is the job that Ravelin does.
The large N represents the proportion of margin wiped out by the cost of chargebacks from one single fraudulent user, or what I term as ‘fraud leverage’. If your business operates on very low margins (for example, marketplaces), the cost of one single chargeback can easily eat up 50x the margin you made from successful transactions, or a fraud leverage ratio of 50x.
Given the dramatic problem that fraud can cause a business, what can be done to solve it? We think that the answer lies in augmenting large scale behavioural data analysis, with the implicit domain knowledge that our clients have about their own businesses.We operate under the motto, ‘if someone’s neck is on the line for your decision, allow them to understand how you came to it.’ Therefore at Ravelin, we make sure our reasons for making a decision are clearly communicated back to a client. We think that since our system influences someone’s performance at work - e.g. a fraud analyst using Ravelin - it wouldn’t be acceptable to provide a black box system that wouldn’t allow them to understand why a machine learning system is making a decision; and more crucially, allow them to improve it. Failure to do this causes frustration and feeling that the system is out of their control. A quote by Kris Hammond illustrates this point rather well: ‘a brilliant machine that throws out an answer to a problem but cannot explain itself will be of little use’.
Another principle is to ‘use tools that make you disproportionately productive’. This may sound like an obvious truism, and to some extent it is, but it’s worth restating the impact that spending the right money on the right tools can have for a technical business. A good example of this is the way we use Google’s BigQuery product at Ravelin. We’ve found it to be robust, extremely scalable and key to the product we’ve built.
Almost every part of our system touches BigQuery in some way, from application logs, traces of request that go through our system and the core event data that our clients send us. It forms the core part of our feature extraction process for machine learning, and allows us to answer any question we wish, about virtually any part of our system. If we had decided to run our own data warehouse infrastructure with the small technical team we have, we’d be distracted from the core proposition of Ravelin: detecting fraud.
In summary, fraud is an extremely fast growing problem. I hope that the above post and talk give you some insight as to the approaches and principles we take as we solve it for our clients at Ravelin.The other talks at the meetup were given by Miriam Redi, a research scientist at Yahoo Labs London, and Nathan Benaich from Playfair Capital. You can watch my talk below. We’re always looking for interesting speakers working in AI, so get in touch if you’re interested in knowing more!