Hands-On Unsupervised Learning with Python

Mini-batch K-means

This algorithm is an extension of standard K-means but, as centroids cannot be computed with all samples, it's necessary to include an additional step that is responsible for reassigning the samples when an existing cluster is no longer valid. In particular, instead of computing global means, mini-batch K-means works with streaming averages. Once a batch is received, the algorithm computes a partial mean and determines the position of the centroids. However, not all clusters will have the same number of assignments, so the algorithm must decide whether to wait or to reassign the samples.

This concept can be immediately understood by considering a very inefficient streaming process that starts sending all samples belonging ...