Addition of kmeanscrossvalidator #61
Conversation
|
Maybe it's a bit nitpicky, but currently this implementation is fully built on KMeans whereas any clusteringmethod should work right? Maybe something for the future to make the CV take any clusterer as a parameter |
|
@MBrouns I agree. Related: I think |
|
@kayhoogland whats your opinion on this? it feels weird to merge this knowing that we will add a more custom thing as well. (also ... typically clustering algorithms are very prone to overfitting on a column if there's no standardisation so i think we don't just want to accept multiple clustering algorithms but that we want to allow for pipelines) |
|
I agree, it seems the logical thing to do. Also, the name is indeed catchy ;) |
|
I did some changes to enable for more clustering methods. A difficult thing to take into account is the n_splits from _BaseKFold. Some (most) clustering methods (for example DBScan) generate splits based on the data itself. |
|
closed it because i hadnt heard from it in a while. ill gladly re-open if there's more attention for it. |
|
I just discussed this with @kayhoogland and we see a good way going forward. I'm putting my feedback in a revie |
kayhoogland
force-pushed
the
feature_kmeanscrossvalidator
branch
from
2541602
to
efaef5e
Compare
sklego/model_selection.py
Outdated
| if isinstance(X, pd.DataFrame): | ||
| X = X.values | ||
|
|
||
| clusters = self.cluster_method.fit_predict(X) |
There was a problem hiding this comment.
For not refitting if self.cluster_method is already fitted:
from sklearn.exceptions import NotFittedError
try:
clusters = self.cluster_method.predict(X)
except NotFittedError:
clusters = self.cluster_method.fit_predict(X)
| super(KlusterFoldValidation, self).__init__(n_splits=3, | ||
| shuffle=False, | ||
| random_state=random_state) | ||
|
|
There was a problem hiding this comment.
It's probably nice to set self.n_splits = None here
|
w00t. |
First version of the kmeans cross validator discussed here #5