I hear it time and time again: “All this data science jargon confuses the hell out of me!”
Well my friend – today I’m going to dynamite the mountain of jargon standing between you and understanding what data science is all about!
While the mountain is a large one, we have the technology to reduce it to rubble.
To begin, let’s look at the three main types of machine learning – supervised, unsupervised, and reinforcement – and some of the jargon you’ll encounter.
Supervised learning uses labeled training data to train a machine learning model which can predict an outcome.
Labeled training data means data that contains the outcome you want to predict. Housing prices, categories, whether someone purchased a product or not – are all examples of outcomes.
Labeled training data contains two things: a target and one or more features. The target is the outcome you want to predict. The features are attributes of the data used to predict the outcome.
Note: I’ve seen outcome, label and target all used to describe the same thing.
There are two types of supervised learning methods: regression and classification.
Regression is used to predict quantities such as housing prices. These algorithms focus on the relationship between two types of variables: dependent and independent. A dependent variable is your target variable – the outcome you want to predict. The independent variables are all the rest – they are used to predict the change in the dependent variable.
The Boston Housing Price dataset in the scikit-learn Python library is an example of a dataset used for a regression:
Here the ‘target’ column is the dependent variable, or the target – it’s what we want to later predict. The rest of the columns are the features, or independent variables.
Classification is used to predict categories such as “spam” or “not spam”. These algorithms are used to assign two or more discrete classes, or categories.
The Wisconsin Breast Cancer dataset from scikit-learn is an example of a dataset used for classification:
In this dataset the ‘target’ column is what we want to predict, and there are two categories – benign and malignant. All other columns are the features.
The biggest difference between supervised and unsupervised learning is that with unsupervised learning, we do not have labeled training data, we just have data. The key here is we don’t know what the outcome should be.
Because we don’t know what the outcome should be, we ask a different question – are there patterns in the data, and if so, what are they? Unsupervised learning algorithms then look to see if there are in fact patterns.
The most common method of unsupervised learning is clustering.
Clustering is essentially a way to group common items. Each group is called a cluster.
The idea is that items in one cluster are like each other, and different from items in the other clusters.
In the following image we have three clusters represented by blue, black and gray circles:
The third type of machine learning is reinforcement learning. While the term may not be familiar to you, examples have been in the news.
The most recent is Alpha Go. In March of 2016, Alpha Go became the first computer program to beat a Go world champion. And it was seen by a few hundred thousand people.
How did it do it? Reinforcement learning.
In reinforcement learning a computer program attempts to maximize some reward. In the case of Alpha Go, the reward is to win a game of Go. In another example, Google saved millions of dollars in data center costs by having a program monitor and maintain the environmental control systems.
How does a computer program learn to play a game? Much like a human, only much faster.
To train Alpha Go, the team first trained the system on thousands of games – each move made and the outcome of the game. Next, they had Alpha Go play against itself over and over again, thousands of times.
The key here is that the program learns less by data provided to it and more by interacting with it’s environment. It’s the equivalent of a parent teaching a child what a dog is, and that child learning through trial and error to label all the other dogs she sees.
Let’s Keep Knocking Down the Mountain!
In this post we started to knock down the mountain of data science jargon by looking at the three main types of machine learning – supervised, unsupervised, and reinforcement.
We looked at what each of these are, what they’re used for, and examples of data used to train the algorithms.
This is just one part of the mountain. We’ll be attacking others.
What jargon have you encountered?