Have you ever tried to find a needle in the haystack? That’s what fraud prevention feels like for most, and this is why fraud detection vendors leveraging machine learning have been burgeoning these past years. After all, it is, to date, the best way to identify issues in the millions of transactions that go through an e-commerce platform daily.
Yet, the use of AI at scale to stop fraud has its own set of challenges: from the hyperdynamic market conditions and fraud patterns, to the use of the same models to cover multiple industries/geographies/merchants, and to the highly imbalanced nature of the predicted event.
To face these challenges, fraud detection vendors need to have a laser-sharp strategy for their ML production phase and make sure to empower their data science and fraud teams. Doing so requires access to insights about the performance of the models in production, and a granular understanding of the predictions, to detect anomalies before it is too late.
In this blog, we look at how fraud detection solution vendors can leverage their ML monitoring solutions to boost the efficiency of their fraud and data science teams and deliver better service to their merchants.
The fraud analyst team’s dilemma
“AI has helped us detect fraud, but there is still a long way to go until we can feel in control of the predictions.” – Luca B, Fraud Analyst, Top US Fraud detection vendor.
As subject matter experts, fraud analysts are the best equipped to detect suspicious behaviors and raise alerts. But as they support a growing number of merchants across verticals and countries, they can’t operate at scale without AI, which exposes them to a “black box” paradigm as they lack the visibility necessary to trust models and predictions.
Fraud analyst teams (all too) often find themselves drowning in anomalous and suspicious activity alerts. They are constantly playing firefighter, trying to manually validate the decisions automatically taken by the AI to avoid merchants’ and end customers’ dissatisfaction – or worse, $$ loss and brand damage.
To make their tasks more efficient they need to have a solution that helps them gain visibility and tame the velocity and quantity of data that goes through their systems every day. It is only once they manage to have a more granular and rapid view of their models’ activities, and of the changes in the data feed behavior, that they can understand the impact of these on the model’s approval rate. And only then can they feel empowered by the AI, and in control of the predictions.
For instance, one of the fraud teams we support managed to dramatically reduce the time it took them to detect a new fraud pattern in a specific industry – from roughly 3 weeks to 3 days, by getting a timely alert on specific segments that contain a high level of uncertainty, and by understanding the data changes that caused it.
In the absence of a dedicated AI Assurance for a model monitoring solution, the fraud analysts team would usually invest a lot of effort investigating the data based on guesstimations. In turn, this would negatively impact their efficiency and that of the business as a whole to approve/reject transactions.
Data science teams: The need for an “efficiency boost” to be ubiquitous
The efficiency dilemma also impacts the data science teams, as analysts would turn to them to get answers about the predictions, assuming that ensuring models work as intended, is a data scientist’s job. But if the data science team is busy troubleshooting the models, then who is in charge of creating the new ones? How efficient is a data science team that needs to be everywhere: researching, training, tweaking, monitoring, and babysitting the models?
In truth, many data science teams still find themselves spending far too much time troubleshooting and manually checking their models once in production, partly due to the fact that the monitoring step is too often overlooked in their overall ML platform strategy and the critical mass of the ML engineering and data science teams’ times focus on the data preparation and model validation phases.
Unfortunately, overlooking monitoring does not only negatively impact their ability to have a timely understanding of the health of the models, it also creates silos and inefficiencies within the enterprise. After all, they are the ones being regularly pulled in meetings to help analysts understand what’s happening with the predictions…
Empowering the data science teams is about making sure they don’t need to spend too much time on the steps described above and have the right insights to optimize their model’s performance, at the right time.
For the data science teams we support, this means being alerted on concept drifts before they become a liability, or understanding when it is worth creating a model for specific merchants. Last but not least, this requires them to have a clear view of when to retrain their models, and with which data to do so while avoiding unnecessary “noise”.
Creating a common language across the organization
At Superwise, we have helped dozens of fraud analysts and data scientists at leading fraud detection vendors, get the granularity they need to trust their models and create better ones.
For the fraud team to observe shifts in transaction requests patterns, and approval confidence at the sub-group level, and for the data science team to have clearer KPIs and prompt weak spots identifications.
With our AI assurance solution, we monitor and assure the health of models in production while alerting when something goes wrong. At the right time. Thanks to our unique approach, we can detect and alert when anomalies occur, so you don’t have to hear it from your merchants when it’s already too late. For some of our customers, this means, reducing the time it takes them to detect, debug/retrain and fix issues from months to a matter of days, and saving them thousands of dollars.