Payplug, the 1st French fintech, launches an innovative programme using Machine Learning to bring better support to merchants, concerning card fraud prevention. The programme aims at reducing the 3-D Secure triggering rate, while confining chargebacks and maximising conversion. Let’s look back at this project proposed by 3 developers, and conceived during R&D workshops with Jérémy Cohen-Solal, Chief Development Officer at Payplug (ex Dalenys).
In which context have you developed this project?
Jérémy: The company dedicates Fridays to presentations, training and workshops for developers. A special time of the week called “Freedev”, to put aside the common roadmap, and allow the teams to step back and stay at the cutting edge of innovation. Since last year, sessions called “Ultimate Friday Coding” have been added to this programme: organized in teams of 2 or 3 developers, everyone can imagine and achieve an experiment to optimise the platform or the service rendered to our clients.
One of these teams, originally made up of 3 developers, has chosen to test Machine Learning on a key topic for merchants: online and offline fraud.
And what is Machine Learning exactly?
Jérémy: Here are a few notions to better understand the concept:
- A model is a mathematical function enabling autonomous decision-making (without any human intervention in the last version of the model), on the basis of several provided parameters.
- A feature is a piece of information provided to the neural network. The features’ selection is crucial and has a major impact on the network’s final efficiency. The main benefit of the network: highlighting relationships between the provided features.
- The training is a phase when we confront the network to historical data, in order to adjust the different weighting of neurons elements, until we converge towards an optimal result.
What was the scientific approach?
Jérémy: After a state-of-the-art research of technologies and existing knowledge on that subject, the team chose to investigate on the neural network’s lead. They then defined their scientific approach, which relies on several key steps:
- Scope of the Proof of concept
Many libraries were considered and tested. But, in an educational perspective (and also recreational, we have to admit!), the team chose to develop its own implementation in C++ and OpenCL to make use of the GPU (Graphics Processing Unit).
Once the dataset built up, the team addressed the question of the most interesting features for the neural network: about 30 out of 100 features available were selected for the Proof of Concept. More features would have made the analysis more complex, without increasing the accuracy.
- Implementation of the first neural network
The developers made the implementation through multiple iterations
- Learning phase of the neural network, on a sample of 6 month of activity
The code was adjusted several times. Repeatedly, the learning had to be started over again from scratch, to get satisfying results (improved sampling, addition or withdrawal of features in input…)
- Confrontation of the neural network to the next 6 months of traffic
During this 6-month period, we knew the proven or theoretical fraud cases, which enabled us to validate the accuracy of the neural network’s decisions.
The first results were analysed, then after a final set of results… the project was mature enough to be presented to the rest of the team.
How has this initiative become a company-wide project?
Jérémy: During this presentation, Dalenys’s CIO – Chief Information Officer – Romain Pera immediately detected the proposition’s accuracy, and a big opportunity for merchants. This Machine Learning programme is a step forward in the fight against fraud: it allows transaction analysis automation, without any prior rule definition.
The decision was quickly taken to scale the project up: the team suggested to integrate the data division to refine the process (features, datasets…), and define a method for comparing and validating the results reached by the neural network. The data team’s strategy to assess the results’ accuracy was mainly based on 3-DS triggering and chargebacks reception.
Payplug’s Machine Learning: the Proof of concept’s main results
In this test, the 3-D Secure trigger volume was divided by 4, while reducing chargebacks’ volume by 4 in some traffic classifications.
What’s the follow-up of this initiative?
Jérémy: Thanks to these excellent results, and in order to build on this initiative, a full-scale test of the neural network is now in the Payplug roadmap, so we can industrialise a service and offer it to all of our merchants. From a small project led by 3 developers, we have now come to a programme supported by the whole company, with impressive results. The Ultimate Friday Coding has been the trigger, and we’ll repeat it of course!
Another lesson: on that kind of innovative and mainly “data centric” projects, as developers we desperately need the confrontation with data scientists, whose scientific and statistic culture, as well as working methods, are a real inspiration.
Ultimate Friday Coding: a real change in your process, or a continuity?
Jérémy: A continuity of course! Innovation is in Payplug’s DNA since its creation. In 2012, we already were the first payment solution to industrialise a tool for merchants to trigger 3DS with a selective approach (“Smart 3DS”). In a rule engine, banking and merchants’ data are analysed all together: cart content, client segmentation, velocity (several transactions conducted over a short time period), several cards used with the same email… this risk-based approach has proven its worth with an average 3DS triggering rate of 15%, and a fraud rate of 0.2%!
Thus, Payplug has always optimised the ratio between fraud prevention and conversion, to enhance merchants’ profitability.
Learn more about 3-D Secure, discover our article: 5 common beliefs about fraud fighting and 3-D Secure.
*Frenchweb 500 ranking 2016 and 2017
