Java: Data Science Made Easy

Choosing K in K-Means

K-means has a drawback: we need to specify the number of clusters K. Sometimes K can be known from the domain problem we are trying to solve. For example, if we know that there are 10 types of clients, we probably want to look for 10 clusters.

However, often we do not have this kind of domain knowledge. In situations like this, we can use a method often referred as the elbow method:

• Try different values of K, record the distortion for each
• Plot the distortion for each K
• Try to spot the elbow, the part of the graph where the error stops dropping rapidly and starts decreasing slowly

You can do it in the following way:

PrintWriter out = new PrintWriter("distortion.txt"); for (int k = 3; k < 50; k++) {  int maxIter = ...