Fix Chinese Whispers clustering to process all nodes each iteration

[SamareshSingh]

Summary

Fixed a critical bug in the Chinese Whispers clustering algorithm where vectors within the distance threshold were not being grouped together correctly.

The Problem

The algorithm was supposed to perform num_iterations complete passes over all nodes in the graph, but instead it was using random node selection. This meant some nodes could be skipped entirely during an iteration, breaking the label propagation required for correct clustering.

For example, vectors with a distance of 0.371814 (below the 0.38 threshold) were ending up in different clusters when they should have been grouped together.

The Solution

Changed the algorithm to guarantee that every node is processed at least once per iteration using a Fisher-Yates shuffle approach:

• Each iteration shuffles all node indices
• Then processes each node sequentially in the shuffled order
• This ensures complete label propagation through the graph while maintaining randomization

[davisking]

Thanks but this isn't what the algorithm is supposed to be doing. I.e. the way it's written in dlib isn't a bug. What is here in this PR is a different, but related algorithm.

[davisking]

Although I agree this is what the original chinese whispers paper said to do. I forget at this point why the version in dlib deviates from that paper, but what's in dlib works really well for the applications it's used for so I don't want to go changing it. Might not be as good for existing users.

[dlib-issue-bot]

Warning: this issue has been inactive for 35 days and will be automatically closed on 2026-03-22 if there is no further activity.

If you are waiting for a response but haven't received one it's possible your question is somehow inappropriate. E.g. it is off topic, you didn't follow the issue submission instructions, or your question is easily answerable by reading the FAQ, dlib's official compilation instructions, dlib's API documentation, or a Google search.