Add size check before trying to split for GMeans

[pberba]

This adds a simple check before trying to split the current cluster into two. Since we have the minimum_samples_per_cluster parameter, there is no point in trying to split a cluster with less than self.minimum_samples_per_cluster*2 samples.

This also prevents the algorithm from running into the state where it is trying to split a cluster with just one sample, which could happen if there are some outliers in your data.

You can recreate this problem using the sample data from HDBSCAN

from autosklearn.metalearning.metalearning.clustering.gmeans import GMeans
import numpy as np

data = np.load('clusterable_data.npy')
# Add outliers
data[0][0] = -1
data[0][1] = -1

gmeans = GMeans()
clusters = gmeans.fit_predict(data)

And this may result to the following error:

ValueError                                Traceback (most recent call last)
<ipython-input-10-e61faf5588d4> in <module>
      1 gmeans = GMeans()
----> 2 clusters = gmeans.fit_predict(data)
      3 clusters

<ipython-input-9-5b649f70193c> in fit_predict(self, X)
     82 
     83     def fit_predict(self, X):
---> 84         self.fit(X)
     85         predictions = self.KMeans.predict(X)
     86         return predictions

<ipython-input-9-5b649f70193c> in fit(self, X)
     37                                                          n_init=self.n_init,
     38                                                          random_state=self.random_state)
---> 39                         predictions = KMeans_.fit_predict(X_)
     40                         bins = np.bincount(predictions)
     41                         minimum = np.min(bins)

~/anaconda3/envs/tm/lib/python3.7/site-packages/sklearn/cluster/k_means_.py in fit_predict(self, X, y, sample_weight)
    996             Index of the cluster each sample belongs to.
    997         """
--> 998         return self.fit(X, sample_weight=sample_weight).labels_
    999 
   1000     def fit_transform(self, X, y=None, sample_weight=None):

~/anaconda3/envs/tm/lib/python3.7/site-packages/sklearn/cluster/k_means_.py in fit(self, X, y, sample_weight)
    970                 tol=self.tol, random_state=random_state, copy_x=self.copy_x,
    971                 n_jobs=self.n_jobs, algorithm=self.algorithm,
--> 972                 return_n_iter=True)
    973         return self
    974 

~/anaconda3/envs/tm/lib/python3.7/site-packages/sklearn/cluster/k_means_.py in k_means(X, n_clusters, sample_weight, init, precompute_distances, n_init, max_iter, verbose, tol, random_state, copy_x, n_jobs, algorithm, return_n_iter)
    314     if _num_samples(X) < n_clusters:
    315         raise ValueError("n_samples=%d should be >= n_clusters=%d" % (
--> 316             _num_samples(X), n_clusters))
    317 
    318     tol = _tolerance(X, tol)

ValueError: n_samples=1 should be >= n_clusters=2

[franchuterivera]

Hello, thanks a lot for the PR!

I think this makes sense in the extreme cases as you point out. I am rebasing the changes to the new code.

[codecov]

Codecov Report

Merging #732 (66ed20c) into master (a10b384) will decrease coverage by 0.11%.
The diff coverage is 62.50%.

@@            Coverage Diff             @@
##           master     #732      +/-   ##
==========================================
- Coverage   85.59%   85.48%   -0.12%     
==========================================
  Files         130      130              
  Lines       10365    10375      +10     
==========================================
- Hits         8872     8869       -3     
- Misses       1493     1506      +13     

Impacted Files	Coverage Δ
...arn/metalearning/metalearning/clustering/gmeans.py	0.00% <0.00%> (ø)
autosklearn/experimental/askl2.py	80.19% <75.00%> (-1.72%)	⬇️
autosklearn/__version__.py	100.00% <100.00%> (ø)
...eline/components/feature_preprocessing/fast_ica.py	91.30% <0.00%> (-6.53%)	⬇️
...mponents/feature_preprocessing/nystroem_sampler.py	85.29% <0.00%> (-5.89%)	⬇️
...ine/components/feature_preprocessing/kernel_pca.py	94.73% <0.00%> (-1.76%)	⬇️
..._preprocessing/select_percentile_classification.py	87.93% <0.00%> (-1.73%)	⬇️
autosklearn/ensemble_builder.py	76.65% <0.00%> (-0.41%)	⬇️
autosklearn/util/backend.py	76.15% <0.00%> (+1.42%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update a10b384...e9ecfa0. Read the comment docs.

[mfeurer]

Thanks a lot @pberba for this contribution and @franchuterivera for rebasing this!