ai Machine Learning

Machine Learning is best suited to specific problems for which relevant, accurate and timely data is available, and where any predictions or insight can be easily acted on. However, the most valuable part of the puzzle is a thorough understanding of the problem to be solved.

This starts with a well-defined objective. We work with our clients to, amongst other things, reduce false positive rates; increase the amount of fraud detected; respond more quickly to new and emerging fraud; and reduce analyst time spent designing, implementing and managing business rules. The first step to successfully deploying Machine Learning is to help our clients understand their primary (and also often secondary) goal(s).

Having identified the objective, the next challenge is to understand the available data. First, we examine the context under which the data has been generated (and how it will continue to be generated). This enables us to identify any biases that need to be accounted for when building models. Next, we build a semantic structure that can expose the clearest signals to our machine learning algorithms. This involves defining and linking entities that are encoded within the data and enables high performing features to be engineered.

Finally, it’s important to understand the environment into which the Machine Learning system will be deployed. Such technology can only make a significant impact if it is placed within an environment where processes are in place that can leverage its intelligence. For example, ensuring accurate and timely fraud reporting leads to the best performing models. How do these processes work within your organisation and is there anything that can be done to improve them to maximise impact?

Investing the time to understand all aspects of the problem faced by our clients makes the rest of our process straightforward and invariably leads to the best results.

Machine Learning for fraud detection fails without a well-designed set of features engineered from raw payment data. We use our extensive experience in developing features that work well for fraud detection to tailor our Machine Learning to your incoming data feeds. This involves identifying novel sources of data that might be incorporated as well as reviewing the importance of any existing features via our automated machine learning configuration process. Undertaking an initial configuration to engineer such features means that fast results can be repeatedly obtained in the future.

However, data enrichment does not end once our models or solutions have been deployed. Our data science team is constantly on the lookout for new features. This feedback process means we are continuously improving model performance and striving to increase detection rates as we respond to new types of fraud.

Different algorithms have characteristics that may perform better or worse for a particular problem. Our suite of machine learning tools provides the most appropriate tool for the problem at hand.

Our optimisation process for making sure the very best models are in production uses hyper-parameter tuning; transfer learning in low fraud environments to catch fraud without historic data; and continuous monitoring for problems such as concept drift.

It is essential to ensure Machine Learning models provide sustained predictive performance against new incoming data. We test our models before deployment using out of sample test data to simulate production performance and generate a variety of performance metrics. Our metrics include raw model performance (mean squared error, f-scores, etc.) and real-world KPIs (money saved, number of alerts generated, number of customers insulted, etc.). These metrics inform our clients of the expected performance of the model and, depending on their risk appetite, leads to model deployment at a particular risk threshold. We then continuously monitor model performance once a model has been deployed and automatically generate recommendations for model retirement, re-training, or new data features to be included.

We also make extensive use of our bespoke fraud strategy tools. Our Automated EDA tool assists with the model creation process by ensuring that models are put in place to capture new and emerging fraud trends. Meanwhile, Impact Analysis provides operational insight into the combined performance of multiple Machine Learning models by simulating production KPIs for review by fraud managers.

In some cases, Machine Learning cannot wholly substitute the knowledge and understanding of a problem that comes from being there day in day out. Our approach to an overall fraud prevention strategy is to utilise the benefits of Machine Learning in close collaboration with the expertise of human analysts. Our solutions work alongside these analysts, saving them time to focus on the emergence of new fraud trends and ensuring that models are quickly put in place to mitigate any losses. We firmly believe in the role of human expertise. Our approach is ‘man plus machine’, rather than ‘man versus machine’.