[MATH-584] KMeansPlusPlusClusterer incorrectly selects initial cluster centers and is unnecessarily slow

• Type: Bug
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 2.2
• Fix Version/s: 3.0
• Labels: cluster clustering

Environment

All environments

• Assignee: Unassigned
• Reporter: Randall Scarberry
• Created: 01/Jun/11 16:50
• Updated: 24/Mar/12 16:16
• Resolved: 05/Jun/11 16:28

Description

The chooseInitialClusters() method declares sum as an int, when it should be double. It also is quite slow because it contains a lot of unnecessary computation. I'll attached a patch which corrects the problems.

I found the problems while comparing an optimized implementation of KMeans++ I've been working on with the one in commons math.