kmeanstf

k-means algorithms using TensorFlow (uses GPU if available):

• k-means with random initialization (Lloyd, 1982)
• k-means++ (Arthur and Vassilvitskii, 2006)
• tunnel k-means (Fritzke, coming soon)

The k-means algorithm (a.k.a. “Lloyd’s algorithm”) and its variant k-means++ are probably the most well-known approximation algorithms for the k-means problem .

Tunnel k-means is a rather new algorithm which differs from the other two by its ability to perform non-local “tunnel” movements making it possible to escape from poor local minima and to find solutions with very low SSE (summed square error). Below is an illustration of the principle. A detailed motivation and description can be found in (coming soon).