LouvainHierarchy returns more clusters than requested with cut_straight about scikit-network HOT 2 CLOSED

Thanks for your feedback.

This is due to the fact that LouvainHierarchy returns a dendrogram with equal heights. When you do a straight cut on such a dendrogram, you may get a larger number of clusters than expected. I've updated the documentation in the develop branch.

Example

scikit-network version: 0.20.0
Python version: 3.8.3
Operating System: macOS 10.15

import numpy as np
from sknetwork.data import load_netset
from sknetwork.clustering import Louvain
from sknetwork.hierarchy import LouvainHierarchy, cut_straight


graph = load_netset("openflights")
adjacency = graph.adjacency

Clustering

louvain = Louvain()
labels = louvain.fit_transform(adjacency)
len(set(labels))

Output: 35

Hierarchical clustering

louvain_hierarchy = LouvainHierarchy()
dendrogram = louvain_hierarchy.fit_transform(adjacency)
labels = cut_straight(dendrogram, n_clusters=10)
len(set(labels))

Output: 35 (depth 1)

labels = cut_straight(dendrogram, n_clusters=35)
len(set(labels))

Output: 35 (depth 1)

labels = cut_straight(dendrogram, n_clusters=36)
len(set(labels))

Output: 152 (depth 2)

np.unique(dendrogram[:, 2], return_counts=True)

Output: (array([0., 1., 2., 3.]), array([2616, 329, 117, 34]))

from scikit-network.

Thanks!

from scikit-network.