How to make your ML algorithms think like a human
The finance industry is a prime use case for machine learning, thanks to the abundant data sets, access to capital and strong incentive for efficiency and predicting future outcomes. While rule-based workflows are well embedded within the industry, many businesses are now turning to machine learning to automate the algorithm building process, especially when it comes to fintech.
As digital services become more widespread, financial organisations need to move beyond rule-based mechanisms and manual data analysis to ensure compliance, security and customer service. Machine learning is more scalable, flexible and reliable when implemented properly, but requires the right data to deliver actionable insights.
This is especially the case when it comes to making predictions about human behaviour. At a recent Finastra developer meetup, I heard from Ben Houghton, Head of Data Science for Barclays Payments, about his data approach and how he makes his algorithms think like a human.
The Human problem
“When I talk about my job people always assume that I spend all my time building algorithms, writing code and teaching machines how to think,” Ben explained. “Most of my time is actually spent on data, cleaning it and matching it, everything. That’s why we need to be creative with the data we put in.”
To explain the need for the right data, Ben used the example of buying a milkshake. If we want to build an algorithm to predict which people will buy milkshakes, we would have a lot of data we could feed into it. This might include customer demographics of those who have bought milkshakes before, the type of milkshake they bought, the flavour, the price and the quantity of units sold.
These are very logical data points to use but they might not be the most predictive since you’re not looking at the data your customer cares most about. Without knowing what data they use to make their decision, it’s harder for your algorithm to copy their logic.
The customer might be more concerned with the nutritional value of the milkshake, how long it takes to consume and how much packaging or inconvenience is involved. As for demographic information, it could be important to input the employment status of customers to find those with the disposable income for a milkshake.
As Ben remarked, “Getting your data right and choosing the right information to feed in is definitely the most important part of building your algorithm.”
Fintech machine learning in practice
One of the core uses of ML is in cross-selling: introducing new products to customers based on their previous behaviour. Using machine learning, you can look at data from previous cross sells to see who has taken out loans or credit cards in the past, create models and combine that with data on existing product affinity across their user base.
By looking at who previously purchased certain products, as well as their previous purchases, the Barclays Payments algorithm can predict which other customers are the most likely to be successful cross-sell opportunities – useful for marketing campaigns or planning new customer journeys.
Machine learning is also used to predict fraud. By looking at previous fraud attempts, they collect markers that indicate a fraudulent transaction, such as time, location, value and point of sale, to instantly flag suspicious behaviour to customers in real time. Success depends on the ability to match data points to customer activity to build tailored models of what ‘normal’ looks like.
The power of preparation
“We spend probably 15% of our time building algorithms that make generalisations. We spend a lot of time getting our project into shape, working out what the problem looks like, or whether ML is even the right tool for it,” explained Ben. “A huge amount of our time goes on data, extracting it from our systems, understanding what we’re working with and cleaning it to ensure we understand all of it.”
Machines can only think like a human if they’re using the same inputs a human would. Success depends on making the right decisions at the start of the project, narrowing your scope onto the right issues, problems and approaches.
For financial organisations looking to leverage their datasets, the first step will always be to shrink the view to focus in on the most personal elements to help machine learning algorithms to do the same.