Machine Learning: A Gentle Introduction
05/08/2018 by Sean Ankenbruck Machine Learning, Modernization - Analytics
It seems like every day there is an exciting new article on the internet surrounding the mysterious topic of data science and specifically, machine learning. So, what is machine learning?
At the most basic level, machine learning is the ability to train a computer or an algorithm to learn without being explicitly programmed to do so. This results in better predictions from your models, something that we will discuss later in this post.
With all the hype surrounding this concept, you would think it is something rather new, right? Not exactly.
Although this concept has been known by various names over the years, the mathematics behind it has been in development for decades. From the earliest use of electrical circuits to represent neural activity in the brain to the development of the first decision tree models allowing for classification, to applications today that you didn’t even realize are using machine learning, this field is constantly changing and can be extremely useful if you know how to embrace it.
Your next question is probably, how can my organization leverage machine learning techniques to make better decisions? Unfortunately, it is not as simple as clicking a button to reveal all of the answers.
Your business needs are unique and these processes are not a “one size fits all” type of solution. Let’s take a look at a few of the different modeling techniques and compare the applications of each.
Machine learning covers a lot of different techniques but we’ll limit our discussion to just three models – linear regression, decision trees, and neural networks. These three methods each highlight some strengths and weaknesses of three distinct classes of machine learning techniques.
The first class represented by the regression are models that are easily interpretable but inflexible. This means that you can understand a lot about your variable relationships with the target you are trying to predict, but lack the flexibility to fit your model to complex or nuanced relationships. This lack of flexibility typically results in poorer predictive power for the model.
On the flip side, more complex models like neural networks are extremely flexible and can make much more accurate predictions, yet fail to provide insight into individual variable relationships.
In between regression and neural networks falls a class of models known as decision trees. These models are flexible in their ability to classify observations within a dataset with great accuracy, yet they are also easy to interpret and explain to someone who does not have an extensive statistical background.
The following chart illustrates the differences between the interpretability and flexibility of the models mentioned above.
To illustrate this using a real example, let’s say that you work for the marketing department of a large financial institution. The Chief Marketing Officer has asked your department to diagnose why customers leave your company, also referred to as customer churn.
First, you need to figure out the question that you are trying to answer. Are you interested in understanding which customer attributes are most indicative of churn? Maybe you don’t care to know the underlying cause, but just want a model that can tell you which customers are most likely to churn each month.
Maybe you’re looking for a balance between the two because you want a model that’s as accurate as possible but you also need to easily explain to the Chief Marketing Officer which customer attributes are most significant in predicting churn.
Selecting which of the above scenarios you fall into is a great first step in deciding which machine learning technique to apply.