Description
Describe the bug
Hi all,
I'm using the sklearn.cluster.BisectingKMeans to perform a clustering, and it worked for a range of k values, until it failed at k=9 (I don't think the k-value is important though). The issue seems to be that it failed to split a cluster into 2, but got 2 identical centers instead. More details are given below.
Steps/Code to Reproduce
I'm afraid I can't provide a minimal reproducible example without relying on external data. I've uploaded the training data to an empty repo: https://github.com/Xunius/dummy_repo
The only python script in the repo has only a few lines, but I'm assuming the npz data file is at the same folder as the py script. The code to reproduce the bug is:
import numpy as np
from sklearn.cluster import BisectingKMeans
npz = np.load('./bisectkmeans.npz')
km = BisectingKMeans(n_clusters=9, init='random', max_iter=400,
n_init=10, random_state=10,
bisecting_strategy='largest_cluster')
X = npz['data']
weights = npz['weights']
km.fit(X, None, sample_weight=weights)Expected Results
No error is thrown
Actual Results
Below is the Traceback info:
Traceback (most recent call last):
File "bisectkmeans.py", line 12, in <module>
km.fit(X, None, sample_weight=weights)
File "/home/guangzhi/.local/mambaforge/envs/cdat/lib/python3.8/site-packages/sklearn/cluster/_bisect_k_means.py", line 411, in fit
self._bisect(X, x_squared_norms, sample_weight, cluster_to_bisect)
File "/home/guangzhi/.local/mambaforge/envs/cdat/lib/python3.8/site-packages/sklearn/cluster/_bisect_k_means.py", line 342, in _bisect
cluster_to_bisect.split(best_labels, best_centers, scores)
File "/home/guangzhi/.local/mambaforge/envs/cdat/lib/python3.8/site-packages/sklearn/cluster/_bisect_k_means.py", line 45, in split
indices=self.indices[labels == 1], center=centers[1], score=scores[1]
IndexError: index 1 is out of bounds for axis 0 with size 1
Printing np.unique(labels) gives:
array([0], dtype=int32)
so the labels array is a length-6843 1d array with all 0s.
The centers array is a 2x7 array, with 2 identical rows:
(Pdb) centers
array([[-1.04880691, 1.47209268, 0.92988155, 1.16810444, 0.98679651,
-1.4344821 , -0.70694216],
[-1.04880691, 1.47209268, 0.92988155, 1.16810444, 0.98679651,
-1.4344821 , -0.70694216]])
Going up 1 level in the stack and at Line 342 (cluster_to_bisect.split(best_labels, best_centers, scores)), it is noticed that the best_inertia variable is a tiny value:
(Pdb) best_inertia
5.7262640016675594e-30
and the array X is a (6843, 7)` 2d array, with identical rows:
(Pdb) np.ptp(X,axis=0)
array([0., 0., 0., 0., 0., 0., 0.])
I'm not sure what's causing this failure, just in case it helps, I'm using some sample weights, and when it failed, I have some quite small weights:
(Pdb) sample_weight.shape
(6843,)
(Pdb) sample_weight.min()
1.1269358725791912e-09
(Pdb) sample_weight.max()
1.418544396650835e-05
(Pdb) sample_weight.sum()
0.03364984267143782
Versions
System:
python: 3.8.15 | packaged by conda-forge | (default, Nov 22 2022, 08:46:39) [GCC 10.4.0]
executable: /home/guangzhi/.local/mambaforge/envs/cdat/bin/python
machine: Linux-5.15.85-1-MANJARO-x86_64-with-glibc2.10
Python dependencies:
sklearn: 1.2.0
pip: 22.3.1
setuptools: 66.0.0
numpy: 1.21.6
scipy: 1.10.0
Cython: 0.29.33
pandas: 1.5.3
matplotlib: 3.2.2
joblib: 1.2.0
threadpoolctl: 3.1.0
Built with OpenMP: True
threadpoolctl info:
user_api: blas
internal_api: openblas
prefix: libopenblas
filepath: /home/guangzhi/.local/mambaforge/envs/cdat/lib/libopenblasp-r0.3.21.so
version: 0.3.21
threading_layer: pthreads
architecture: Haswell
num_threads: 8
user_api: openmp
internal_api: openmp
prefix: libgomp
filepath: /home/guangzhi/.local/mambaforge/envs/cdat/lib/libgomp.so.1.0.0
version: None
num_threads: 8