Features/1674 optimization of k means initialization

[Hakdag97]

Description

The bottleneck of k-means clustering (concerning runtime) is the initialization of centroids, which was previously built on a cost intensive serial algorithm. The aim of this pull request is to replace this algorithm by the more sophisticated k-means || initialization of centroids.

Issue/s resolved:

• [Bug]: Error in the function nonzero for certain inputs #1746
• Optimization of k-means initialization #1674

Changes proposed:

• Appropriate handling of an additional edge case in the function nonzero
• Complete new implementation of the initialization of centroids used for k-means, k-medians, and k-medoids
• Adjustment of classes (like KMeans) to match with the new implementation

Type of change

• Bug fix
• New feature

Performance

• Reducing the runtime of initialization of clustering algorithm in distributed and non-distributed mode with split=None and split not None by (at least) an order of magnitude (depending on the setting concerning, e.g., size of data and chosen parameters)

Does this change modify the behaviour of other functions? If so, which?

• yes: the classes KMeans, KMedoids, KMedians and the function where are affected

[github-actions]

Thank you for the PR!