Real-world machine learning for debt collection by Hannah Dockings, AI engineer and KTP Associate | Flexys

News & Opinion

Real-world machine learning for debt collection by Hannah Dockings, AI engineer and KTP Associate

by Flexys Marketing Team

What is a KTP?

The Knowledge Transfer Partnership (KTP) scheme is designed to connect businesses that have an innovative idea with additional expertise to help realise its full potential. It is a three-way partnership between a university (in this case The University of the West of England), a business (Flexys) and a KTP Associate who is a suitably qualified graduate (myself) to lead the project.

Academics from UWE’s Computer Science Research Centre are keen to engage with forward-thinking companies such as Flexys to turn the fruits of their research into benefits for businesses, and society in general. Professor Jim Smith commented ”from the outset we have been impressed with Flexys’ blend of enthusiasm and pragmatism in testing and adopting new approaches to improve outcomes for their clients and their customers. The global pandemic of 2020-21, with its knock-on effects for employment and personal finances, has only served to highlight the importance that applications of AI are underpinned by a strong sense of corporate responsibility and ethical behaviour. These are values where we have found common cause with Flexys since the start of our collaboration.”

The opportunity to gain experience in both an industry and research setting was what initially attracted me to the role at Flexys. I was welcomed into an ambitious research and development team who are using artificial intelligence to improve the debt resolution process. I learned that too often, individuals with complex backgrounds and circumstances are put through a “one size fits all” collections process. 

Flexys is making overdue improvements to the status quo with the help of smart technology, machine learning and data science. We have made great progress in using machine learning to help us improve communications strategy. In particular, how to maximise engagement for customers who are in arrears and to match different groups of customers with the messaging they are most responsive to. 

It’s all about the data

One of the key challenges for any organisation that wants to capitalise on machine learning is understanding how to best utilise the available data. Having domain knowledge and understanding the data are key to a successful application of machine learning. This, along with preparing and processing the data, can take time and is a task that should not be taken lightly. Machine learning can’t do all of the work for you. In the debt collection space, your supplier has to understand the nuances of how the existing arrears process works, how customers interact, and how/when this data is logged, so that you don’t make erroneous assumptions that can invalidate anything learned by your algorithms. 

Understanding the context of the data is also crucial. One example of this would be how the coronavirus pandemic impacted the trends that we saw. While machine learning models can be powerful, they will not be able to understand how or why this data is different, so it is up to the engineer to adapt the machine learning process appropriately.

In-house versus out-of-the-box

The collaboration with UWE has been invaluable when working out how to gain actionable insights from data and apply machine learning. In our case, the out-of-the-box solutions available from various machine learning libraries weren’t flexible enough. The artificial intelligence experts from UWE provide the expertise and creativity necessary to produce an appropriate machine learning solution and have helped to shape the research and innovation at Flexys. 

The main output of our research so far has been an explainable model for predicting if and when customers are likely to engage after receiving messages regarding their arrears. We took a probabilistic approach meaning that the degree of certainty that the model has is inherently transparent. 

To explain, most machine learning methods centre around rigid binary predictions, whereas we have centred our system around finding the likelihood of certain outcomes, which gives us a better indication of how different groups of customers behave overall. It means that even if our clients have limited data and we can’t be 90%+ certain about how customers will behave, we can find general trends and patterns which are still very valuable to learn.

Using insights

We are developing a process using this type of model along with trialling different kinds of messaging informed by behavioural science insights, to continually improve aspects of communication to customers, leading to increased engagement and better debt resolution. An intelligent and adaptive system for customer communication is the end goal, and we are on track to achieve this.

The customer at the heart of everything

One of the key objectives of a KTP is to provide the associate with experience and personal development opportunities so that they are equipped for a successful career in their chosen field, as well as to successfully carry out the KTP research. Flexys have embraced this and shown that they value and support the growth of their employees and are dedicated to research and development. As a result, I have learned a great deal and developed my abilities as an artificial intelligence engineer. Additionally, I’ve been given the opportunity to have a sizable impact on the business, which is quite rare for someone early in their career. 

Most rewarding of all, the benefits of the KTP extend not only to myself and Flexys but ultimately to our clients’ organisations and their customers, who are at the heart of the collections process.

If you’d like to hear more about our innovations in debt collection, please

New call-to-action