From the course: Introduction to Artificial Intelligence

K-means clustering

From the course: Introduction to Artificial Intelligence

K-means clustering

“

- Another common machine learning algorithm is K-means clustering. K-means clustering is an unsupervised machine learning algorithm. It's used to create clusters based on what your machine sees in the data. Let's go back to the animal shelter in Chicago. The shelter used to have a large social room where the dogs got together, sniffed and played. The dogs had their group of friends and they played and hung out together. Each time they had a social hour, they would self-organize into these social groups. Now imagine that the shelter was closing and all the dogs were going to be distributed to three different shelters. To make it easier for the dogs, the organizers decided to cluster the dogs based on their friend groups. So the shelter created three clusters. That means the K in K-means equaled three, because they wanted to divide the groups into three clusters. To start, the machine put a red, yellow or blue colored collar on three random dogs. Each color represented a potential cluster based on their social group. These would be your three centroid dogs. Each of the centroid dogs would look at the mean or average distance between itself and all of the surrounding dogs. Then the machine would put the same colored collar on the dogs that were closest to a centroid. Since these centroid dogs were selected randomly there was a pretty good chance that they wouldn't really have any good clusters. Maybe all three centroid dogs were in the same social group. If that happened, then most of the dogs would be far away from their nearest centroid, so it would redistribute collars until the algorithm could find a good centroid dog. The machine would try over and over again until it picked the best centroid and it might even do this one cluster at a time. At the end of each iteration, the machine learning algorithm checked the variance between each dog and the centroid. Once there was a good centroid, then it would put the same color collars on the friends of the dogs in each cluster. Now, keep in mind that the dogs themselves did not cluster into three groups. There might be seven or eight different social groups but there were only three shelters, so the algorithm had to do its best to cluster the dogs' natural social grouping. The algorithm also had to consider very social dogs that jump from group to group. If the dogs were moving around too much, then it'd be difficult to form real clusters. Another challenge with k-means is it can be very sensitive to outliers. So if you had a dog that wasn't hanging out with other dogs it would still have to be clustered into one of these three groups. Essentially, the dog would be forced to find friends. Organizing dogs into three clusters for three different shelters is probably not a problem you'll run into every day. But K-means clustering is one of the most popular machine learning algorithms. One of the more interesting applications is when retailers use clustering to decide who gets promotions. They might have the system create three clusters to find loyal customers, regular customers and lowest price shoppers. Then they'll create strategies to elevate regular customers into loyal customers. Many organizations are looking for better ways to cluster together their customers. If they can get all of their loyal customers into one cluster, then they can really improve their business. K-means clustering is a good way for the system to cluster people or things by looking at hundreds of different variables.

Contents