We’ve written many blog posts on the topic of Machine Learning and how it’s improving everything from fraud prevention, direct marketing in retail and the customer experience in call centres to getting us to make more impulse purchases online and making holiday and business travel more enjoyable. It’s with good reason we’ve given Machine Learning so much focus on our blog: it is a driving force in what the founder of the World Economic Forum – Klaus Schwab – is calling the Fourth Industrial Revolution.
In this blog post, we answer some frequently asked questions about Machine Learning, starting with what it is and how it relates to Artificial Intelligence.
What is Machine Learning?
Rather than redefine, let’s just quote Wikipedia’s definition of Machine Learning which perhaps best explains it: “Machine learning is a subfield of computer science that evolved from the study of pattern recognition and computational learning theory in artificial intelligence. In 1959, Arthur Samuel defined machine learning as a ‘Field of study that gives computers the ability to learn without being explicitly programmed‘.”
Simply put: Machine Learning is the use of algorithms to teach computers to learn and teach themselves by trial and error through the using algorithms and computational power to process vast amounts of data and automatically find patterns and crunching of data
In the past, static predictive models were carefully crafted from offline data. However, with the faster computing power and the plethora of analytical toolsets that are available today, the analytical model building can, to a large extent, be automated. Using algorithms that iteratively learn from data, machine learning allows computers to find hidden insights without being explicitly programmed where to look.
How does Machine Learning relate to Artificial Intelligence?
When we talk about Artificial Intelligence, we are referring to computational systems that behave like a human – i.e. they can understand the written and spoken language, they can converse and interact “like a human”, and they can process complex concepts like sarcasm and humour. In developing this computational system, the computer needs to learn, a lot, and this is where Machine Learning comes in. Machine Learning is therefore a fundamental branch of Artificial Intelligence.
How does Machine Learning relate to Predictive Analytics?
Although most applications involve the requirement to predict something (like a response, a fraudulent application, an attrition account), there is a tranche of Machine Learning that does not predict anything specific. Customer segmentation is a good example, where one looks at the customer base and the requirement is to group the accounts into homogenous clusters, without pointing to any specific target variable. Anything that involves the prediction of something can be considered to be under the wider Predictive Analytics banner.
There are two types of predictive Machine Learning: static (or batch) and dynamic (or incremental). The more stable and trusted approach would be to manually extract the relevant data, train the models with care, validate and then implement the models. With the computing systems and algorithms that are available today, full dynamic Machine Learning is possible which allows for the retraining process to be carried out automatically, with little or no human intervention. A good example of this is the iPhone’s Siri which learns the lifestyle patterns of the user and automatically adjusts its recommendations accordingly. For most Machine Learning projects within the financial services industry, we do not advocate this black box approach, but rather the development of stable and explainable predictive models. Dynamic Machine learning algorithms can then be applied on top of this solid foundation in order to bring in more recent patterns and trends, further lifting the performance of the predictive analytics system.
Why has machine learning become such a hot topic?
The reason for the increasing interest is due to the significant increase in data that is now available, which makes machine learning more relevant, accurate and more effective for more businesses than ever before. Machine Learning automates decisions by analysing large and diverse datasets at lightning speeds, predicting what would lead to a positive outcome and making or taking the recommended action. The accuracy of machine learning depends on the amount and the reliability of the available data. With the data deluge happening now thanks to a combination of smartphones, the Internet of Things, the Internet, RFID technology and social media there is now an unprecedented amount of data being generated every second of the day for machine learning algorithms to analyse, search for and recognise patterns and trends in the data, and to then output a decision or recommended action.
What are the potential benefits of creating machines that can be programmed to learn?
Machine Learning automates decisions in a continuous improvement cycle. And machines automate actions. The combination of both greatly reduces the need for human intervention for a process to be completed. When we remove the reliance on human intervention we tend to achieve faster response times, a reduction in cost and a reduction in human error or bias for any process – from the granting of a loan to the landing of a plane. Some examples of how machine learning is being used are to instantly identify fraudulent behaviour on an account as it happens and automatically send an alert; flagging customers who are most likely to unsubscribe or leave a business for a competitor based on real-time analysis of customer behaviour and transactional data and then take appropriate action (or not as the case may be); adjust pricing on thousands of items at once based on availability and demand; and making real-time recommendations of products or services based on a person’s purchase behaviour as they shop or complete a purchase. Decisions based on machine learning get better or more accurate with time, as they learn what decisions and actions lead to a negative or positive outcome. As the algorithm is fed results of its actions and consequent outcomes it learns and adjusts to do more of X to get the desired result instead of Y. It’s getting closer and closer to Artificial Intelligence every day.
How does Machine Learning relate to Cognitive Computing?
Again, to quote Wikipedia: Cognitive computing is the simulation of human thought processes in a computerized model. Cognitive computing involves self-learning systems that use data mining, pattern recognition and natural language processing to mimic the way the human brain works.
The world is currently generating data more rapidly than we are able to analyse it, is cognitive computing the answer?
Yes. Although the human mind is complex and powerful, we are unable to compete with computers when it comes to the speed at which they are able to analyse large sets of data and recognise patterns and trends. And they are just getting faster and smarter. So, although the human race has been able to achieve some incredible things and there doesn’t seem to be any problems we cannot solve, in some cases, we lack the luxury of time to solve them. Cognitive computing could help us find answers or solutions to these problems faster.
How did you use machine learning to forecast the results of Rugby World Cup matches with a high level of accuracy? How was this possible?
Sports fans make their own predictions on match results based on their own experience and observations, as well as gut feel and personal biases. With machine learning, we based our predictions primarily on data we took from 6,000 matches played by 99 teams since 1995. As we then progressed through the World Cup matches, we incrementally added the results from each match to our model, so the model could adjust its prediction based on what happened in the last few matches. What was interesting was that the model didn’t look at who played the matches, but rather the characteristics of the teams that played the matches. So, there was no bias towards any team – it was just looking at the data or the characteristics of a specific in making its prediction. When we added that new data, the model adjusted and compared the characteristics of the teams whose data was added and the result of these teams’ matches and adjusted its expectations for teams with the same or similar characteristics. We continually added new data and the model continually adjusted to this new data, increasing the accuracy of our prediction with every match played. One interesting set of data we included that added a bit of the human element to the prediction was the fantasy value of the players and the bookie odds. We didn’t look at fields, or influencers, such as referees and weather as this would have taken a lot more time and effort and would not necessarily have resulted in a large enough increase in accuracy of the prediction to justify the added effort.