Game changer: How machine learning technology became easier to deploy

Since our inception machine learning has been a source of equal excitement and frustration for our in-house data scientists’ and analysts’. The team could see the potential of machine learning to improve our products but long established technology barriers prevented us testing our ideas.

Machine learning algorithms have existed since the 1950s when computer scientist and cogitative scientist Professor Jon McCarthy pioneered the concept. But in the past it has been difficult, if not impossible, to run the technology due to the computing power, the depth and volume of data required, and the need to write your own algorithms. With the advent of cloud-based technology and the availability of off-the-shelf algorithms, however, machine learning technology can now be deployed much more easily, especially with the right specialist knowledge, resources and data sets.

Following on from our machine learning whitepaper we want to explain why we think deployment is no longer a barrier to harnessing the power of machine learning and realising the exciting potential it offers. We’ll bring this journey to life through our work with Microsoft and the testing of our own models in the credit and risk space.

Our ability to deploy machine learning models to our own data sets comes down to the fact that computing infrastructure and programmes, such as Microsoft Azure and Amazon Web Services, have become more advanced.

At Callcredit our data scientists use Microsoft’s cloud computing platform, Azure, to deploy machine learning algorithms. Azure offers users vast computing power and enables our data scientists to publish machine learning models as a fully featured web service that is accessible, flexible, dependable and scalable.

We’ve been longstanding partners with Microsoft. They have been integral to our success and they are rightly viewed as leading the way in secure cloud computing by those in financial services. So for us the Azure platform was the logical choice when we were looking to explore how we could use machine learning.

TransUnion (formerly Callcredit) Data scientist Jonathan Gill explains; “As a business with rich data sets Microsoft Azure provides a real opportunity to sandbox machine learning algorithms and to look at ways of improving performance. With it [Azure] it is easy to deploy models and publish them as a web service. Which means you can pass what the model needs [such as] a person’s credit history as a web service and it’ll fly back to you within fractions of a second.

“There is also no need to re-code the models into a separate language, instead you are able to call them immediately…[it’s] an elastic service that can easily scale depending on clicks and offers easy integration for users. They just need to call the web service [API].”

To illustrate the game changing nature of the platform Gill cites how the team made some modifications to a live machine learning algorithm and within a day had redeployed it back into the fraud service it was being used in. This agility is far removed from the old, inherently inflexible world of machine learning where you would need to re-write code, buy hardware if clicks went-up etc. to support a new service or scale a service.

The platform provides TransUnion (formerly Callcredit) with the potential to extract previously hidden value from our data. With Azure’s expansive range of capabilities we can store, catalogue, analyse and transform our data assets. And we can create new rule sets, event triggers and models that enable us to mine the data and extract value in new and innovative ways.

When thinking about our own challenges, such as the way we can enhance our capability to predict credit risk, detect fraud or calculate ‘next best action’, this presents exciting opportunities for new products and the evolution of our ecosystem, which will ultimately benefit clients and consumers.

In 2015 a group of our data scientists visited the Microsoft Research Lab, Washington, United States, giving them a first-hand opportunity to learn about the advances in machine learning. The team followed this up with a visit to the Bristol data centre (UK) where they were blown away by the scale – imagine rows of football pitches full of computers and servers – and the security in place.

For Gary Diplock, a data scientist at Callcredit, working with Microsoft has been productive; “Callcredit and Microsoft have a long-standing working relationship stretching back over 15 years. We have access to their subject matter experts, who help us get up and running more quickly by sharing their considerable experience using and applying their tools and technology. We also look to share feedback on our use of products and services, as well as suggestions for future enhancements.”

Continuing the theme of collaborative learning, Gary currently runs machine learning meet-ups at our Leeds office, that feature a mixture of internal and external speakers, the latter drawn from both industry and academia.

Over the last 24 months our dedicated data science team have conducted machine learning experiments, focused on the complete customer lifecycle, from acquisition through to collections. We’ve sandboxed a number of tests and researched a wide range of techniques and approaches to establish industry best practice.

Our tests have acted as a proof of concept of how machine learning can be used. They demonstrated a significant uplift from deploying machine algorithms over and above traditional scorecard approaches. For instance, we’ve achieved a 5 to 10% Gini point uplift to our online sign-up fraud detection services. The experiment was so successful that we’re now working to develop commercial machine learning scores. Look out for our next article that focuses on how we’ll harness machine learning to give our client’s competitive advantage in 2018.