Checklist - is Machine Learning the right tool?

There is a lot of talk about Machine Learning nowadays and yes, Machine Learning is an incredibly powerful tool, but it doesn't work in all situations, as you might be led to believe when you hear and read how some people express themselves about it. To clear up the confusion, here is a checklist for recognising a problem that can be solved using Machine Learning

1. Question

First, you need to start with a question or problem you want answered. It's easy to start by rummaging around in the data you have, but that risks taking the focus off what's important for you to know and just confusing you even more. So don't start with the data, start with a problem that is important for you to answer.

2. Learning (not automation)

It is important to identify whether learning or automation is required to solve the problem. If there is a series of well-defined steps to be performed and you can formulate rules for each step and each situation - then it is not Machine Learning that is needed, but plain old automation. But if learning is required, that we can't define all possible situations and cases in advance from the beginning - then it might be a case for Machine Learning.

3. Prediction (not causality)

Do you want to know if one leads to the other? Unfortunately, Machine Learning can't help you with that. But if it's enough to identify a correlation and find out statistically that two or more factors tend to be related, Machine Learning can certainly help. Machine Learning can be used to predict whether a particular factor is likely to be met given other factors, for example whether a customer who has bought a particular product is likely to want to buy another product.

5. Reasonableness

Is it reasonable to assume that the solution to your question is in the data you have? Example: a research team at the University of Pittsburgh Medical Center wanted to develop a decision support tool for physicians that would help them decide whether a patient with pneumonia was at higher risk of developing complications and therefore should stay in the hospital, or whether the risk of complications was low and the patient could be cared for at home. The researchers tested different methods, including machine learning to predict risk, and machine learning methods made better judgements than other methods in most cases. But there was one case that got weird: following the computer's advice to the letter meant that patients who had both asthma and pneumonia would be sent home, which didn't make sense. It turned out that the practice at the hospital for patients with both asthma and pneumonia was to send them straight to the intensive care unit, and the care they received there was so good that the risk of complications was very low. However, this did not mean that it was appropriate to send pneumonia patients with asthma home, quite the opposite! However, the machine learning algorithm did exactly what it was supposed to do - based on the data it had available, the risk of complications was actually lower for asthma patients.

6. Intuition

And then the final common-sense check. Does your intuition agree that the data you have can answer the question? Of course, with Machine Learning you can come to new insights and realize new relationships, but it's still not magic. If it feels magical, there's probably a flaw somewhere and you should revise the project.

Although Machine Learning as a concept has been around since the 1950s, it is only recently that it has really come into its own. This is because we now have better computing power, better algorithms and larger data sets. AddPro is exploring Machine Learning and is interested in finding new applications for it together with customers. So if you have a problem you think can be solved using Machine Learning, we'd love to talk to you and see what we can do together. We like to start small with an experiment or a proof of concept to measure the potential and then can scale up as we see the solution generating value.