I recently had the time and opportunity to bone up on machine learning.  The opportunity that presented itself was a 10-week on-line course taught by a well known Stanford Computer Science professor, Andrew Ng.  The course, offered by coursera.org included lectures, quizzes, and programming exercises. It was great. And it really got me thinking about applications for payments companies.

With the plunging cost of computing and the growing need to manage and interpret “big data”, it seems to me that machine learning should be an important ingredient in payments company IT plans.  Of course, there are many large companies in the payment space that already employ machine learning – but for other smaller companies and newer entrants, machine learning may represent an untapped opportunity.  Recent developments in machine learning may also be relevant for the big players.

What follows is a summary of my take-away from the class: an overview of machine learning, how machine learning can be applied to the payments world, and some observations on how to move forward.  I am by no means a technical expert in machine learning (it is a branch of computer science), but know enough after some hands on experience to assess its relevance to a particular application and understand how to approach solution development – in short, machine learning from a management perspective.

Machine Learning Overview

Machine learning was defined by Arthur Samuel as a:  “Field of study that gives computers the ability to learn without being explicitly programmed.”   For machine learning to be effective, there must be data available from which to learn, and it must be possible to learn from the data.

There are two distinct types of machine learning:  supervised learning and unsupervised learning.  Supervised learning occurs when the learning data contains the “right answers” as shown in diagram* below.

Some “Machine Learning Speak” is used in this and diagram* below.  “Features” are simply the data from which machine learning algorithms are created.  Used in supervised learning, “Labels” are the right answers from historical data.  The machine learning algorithm is used to predict results for a new set of input data, the “Expected Label”.  An example of supervised learning would be a spam detector, which had learning data that included emails marked as “Spam” and “Not Spam”.

Unsupervised learning occurs when there are no “labels” – the data must be organized, but there is no right or wrong answer from the historical record. Customer classification into different groups based on multidimensional learning data is one example.

Machine learning can be employed to do any of the following;

• Forecast the sales price of a home
• Provide yes or no answers – i.e. is this spam or is it not?
• Identify anomalies – such as aircraft engines that might be subject to failure
• Recommend books or movies that a user might want to see
• Perform facial recognition
• Read hand-written numbers

Machine Learning and Payments

The opportunities for machine learning in payments are almost limitless.  Some examples are listed below – but many others are possible depending on your position in the payments value chain.

• Transaction risk management.  Use machine learning supervised learning algorithms to identify risk of payment transactions.
• Acquirer or ISO merchant risk analysis.  Use machine learning to assess acquirer risk in signing a new merchant or managing risk of on-going merchant relationships.
• Optimizing user’s web experience.  This is commonly done by the high volume web companies, and may have particular relevance to customer check-out.
• Customer clustering.  Use machine learning unsupervised learning to group customers based on a set of customer characteristics.

I welcome your input on additional payment applications make sense for machine learning!

Moving Forward With Machine Learning

For those firms that are experienced with machine learning, there is nothing new to report.  But if you are just starting out, or have not kept abreast of recent developments, here are a few things to keep in mind:

1. Machine learning can be applied to a wide range of opportunities.  Remember the prerequisites:  data must be available, and there must be a strong likelihood that the data correlates to a solution.
2. There are a number of ways to approach machine learning.  It is important that the path you select is properly balanced between cost and return.
3. When focusing on a particular machine learning solution, there are methods that can be employed to reduce cost of solution development.  For example, it is often assumed that the more data that is available, the better the machine learning solution will be – but this is not always the case.  There are tools available to help manage the machine learning development in a way that is resource effective.
4. Large scale machine learning systems can benefit from parallel processing (attacking a single problem with multiple computers).  Cloud services such as Amazon Web Services offer opportunity to bring multiple processors to bear on machine learning algorithms.

Feel free to reach out if you’d like to discuss the implications of machine learning for your payments business.

* Diagrams courtesy of Scikit-learn: Machine Learning in Python, Pedregosa et al., JMLR 12, pp. 2825-2830, 2011.

[Comment from Scott Loftesness: I recently came across and read Eric Siegel’s “Predictive Analytics: The Power to Predict Who Will Click, Buy, Lie, or Die” – it’s an easy read that provides a great overview of machine learning techniques and some applications!]