August 20, 2020
A Sigh of Relief: Learning How to Not Turn Away Legitimate Customers with Machine Learning and Data Science
By Byron King, Business Data Scientist and eCommerce Vertical Owner -- Nethone
Brick-and-mortar stores have gone to great lengths to protect their merchandise from theft. They have a number of theft-prevention tools at their disposal, but sometimes they still get it wrong. There is a famous scene in the film Pretty Woman in which the main character, Vivian Ward, goes shopping in a high-end Rodeo Drive boutique in Beverly Hills, only to be turned away by the employees because she does not look the part of their typical customer. It is a nice scene in a Hollywood movie, but the fact is that the same situation happened to several people I know personally, and there are even examples of some of the richest people in the world being turned away from brick-and-mortar stores because they did not "look the part."
Now imagine that you are the owner of a high-end fashion/jewelry eCommerce platform and you are told that you lost a legitimate customer that wanted to buy a $30,000 handbag, but she was turned away because of a legacy system that generated a false positive based on a simplistic review. Does it not make your heart hurt?
The good news is that there are an ever-improving array of tools available to you today to ensure that the occurrence of such a situation is minimized.
Our company recently worked with a firm that experienced the pain of discouraging real customers with ever-increasing manual reviews of transactions. We felt their pain and did our best to give them the smartest tools available to increase conversion rates and lower both their chargeback rate as well as the amount of manually-reviewed traffic.
In short, what this eCommerce fashion giant faced is more intense and globalized than other industries, but actually the same type of problems as what many other organizations face:
- The continuous growth of manual review of traffic, which was generated by legacy anti-fraud solutions providers;
- High chargeback ratios -- the company set a bold KPI to keep them below 0,5% for all of its brands;
- Weak chargeback risk detection, which translated into growing financial losses
Machine learning models depend on the quality of data they are fedWe created tech that sits in the client's environment, gathers myriad data points, and passes it on to the machine learning models which give actionable recommendations. We look at 5000+ data points, most of them non-declarative, proprietary data extraction methods regarding the user's hardware, software, network, and behavioral contexts, both web and native mobile; plus proprietary device marking techniques, 42 configuration based and 8 storage based. It is one-of-a-kind raw behavior understanding analytics necessary for hyper-personalized analysis in both web and native mobile contexts. It sets us apart from our competitors in the space (feel free to ask around).
Learn from past mistakesBut the good thing about coming into a project where there is a history of mistakes is the ability to learn from them, especially if you have a record of the mistakes made by previous systems! So Nethone took an iterative approach to model building. New models are built that target the incorrect predictions of previous model builds.
To detect chargebacks, the Nethone Data Science team implemented two supervised machine learning models. The goal of the first one was to detect chargebacks directly, while the second one was to replicate the results of the decision process of previous anti-fraud providers. Then, the DS team trained and tested multiple machine learning models using various statistical techniques. We found the best performing models and put them to work.
Final touchesLuxury goods eCommerce, like all online consumer goods platforms, see a sharp spike in purchases during the holiday season. The team scaled its data processing resources to ensure both models' predictions would not experience any downtime during important sales seasons, such as Black Friday or the winter holiday season.
Finally, Nethone has already successfully worked in so-called "risky geographies" and we have learned the attributes of real customers versus fraudsters and friendly fraudsters. So we passed on the knowledge to our client's tailored machine learning model. So the client continues to benefit from lessons learned from the experiences of our other clients.
In short, we learned from the past, tested multiple ML models in parallel, scaled for big events, optimized, opened up the client to new markets that were deemed risky, and optimized some more. We were able to cut down the platform's manual review rate by nearly 60%. At the same time, the percentage of accepted traffic increased in comparison to the rates of other anti-fraud solutions providers. The overall chargeback rate decreased by over 12%. For one of the fashion eCommerce platform's most valuable brands, the chargeback rate was lowered by as much as 89%. The company's internal KPIs were met -- the overall chargeback rate was kept below 0.5% for all of its brands. And all of it happened within four months of full integration.
Their fraud management team got to take a breather, too, as they had been burdened with growing manual review of transactions!
Caption: Reduction in manual review and chargebacks means lower costs for the company. Transactions with reduced friction improve the customer experience, which translates into customer loyalty and return purchases.
Nethone is a global provider of AI-driven KYU (Know Your Users) solutions that allows online merchants to understand their end-users and prevent online fraud. By using machine learning technology, Nethone is able to detect and prevent card-not-present fraud, including protection against account takeover. Founded in 2016 by data scientists, security experts, and business executives, Nethone successfully cooperates with eCommerce, digital goods, travel, and financial industries on a global scale.