[ENH] Louvain Clustering: Add cosine similarity

[janezd]

As suggested by @pavlin-policar in a comment to #4855, cosine distance can be computed as Euclidean on row-normalized data, because we care only about ranking. This allows NearestNeighbours to use ball trees.

Includes

• Code changes
• Tests
• Documentation

[janezd]

@pavlin-policar, I didn't really read the entire widget's code, but it seems that it would suffice to normalize the data and change the metric in matrix_to_knn_graph. As I see, the absolute distances are not needed at all (return_distance=False, and I see no other mention of metric).

Am I correct?

[codecov]

Codecov Report

Merging #4864 into master will decrease coverage by 0.00%.
The diff coverage is 91.07%.

@@            Coverage Diff             @@
##           master    #4864      +/-   ##
==========================================
- Coverage   84.17%   84.17%   -0.01%     
==========================================
  Files         277      277              
  Lines       56541    56503      -38     
==========================================
- Hits        47596    47564      -32     
+ Misses       8945     8939       -6     

[pavlin-policar]

As I see, the absolute distances are not needed at all

Yes, that's right. We build a graph, and the edges are weighted with the Jaccard coefficient of shared nearest neighbors between points. Distances are not used anywhere.

This codecov is doing strange things again...