[BUG] Bug in ClusterCentroids estimator

[lixiilu]

Dear author,

We found a bug about ClusterCentroids when we want to under_sampling. The bug is when we input the positive ( minority class ) and negative (majority class ) samples, after ClusterCentroids to under_samping the majority class, we found that there have minority class minority in the final result. Looking forward to your reply.

[chkoar]

Could you post a minimal reproducible example?

[lixiilu]

file: imbalanced-learn/imblearn/under_sampling/_prototype_generation/_cluster_centroids.py
line: 174
X, y, self.estimator_.cluster_centers_, target_class should be X[y == target_class], y[y == target_class], self.estimator_.cluster_centers_, target_class

[chkoar]

Well, that was not a minimal reproducible example, but anyway.

You noticed that because the default strategy is to locate the nearest neighbors (hard) of the centroids and not the centroids themselves (soft).

Since we perform clustering in minority class it make sense to search for neighbors only from the minority class.

Thanks for the report.

[chkoar]

You could raise a PR if you want to. This needs a test also.

[lixiilu]

The strategy we use is to locate the nearest neighbors (hard) of the centroids. And we perform clustering (under sampling) in the majority class samples, however, we find minority class samples in the result of under sampling.