Hi Sam,
I agree that a meta estimator would be the way to go. However, the question is:
are these techniques used widely enough to warrant inclusion in scikit-learn? I
have the impression that they are used much less than agglomerative approaches.
I have myself used them in the past with succe
Andreas,
That's pretty much what my idea for this would be. By default, the
algorithm uses bisecting kmeans but you can specify any clusterer that
follows the scikit-learn api or any function that follows a specific API. I
think that there are some interesting possibilities with allowing the
clust
This feels a bit like it should be a meta-estimator using an arbitrary
clustering algorithm to create a divisive one.
That would easily allow the PCA thing.
On 05/17/2015 07:44 PM, Joel Nothman wrote:
Hi Sam,
I think this could be interesting. You could allow for learning
parameters on each
Hi Sam,
I think this could be interesting. You could allow for learning parameters
on each sub-cluster by accepting a transformer as a parameter, then using
sample = sklearn.base.clone(transformer).fit_transform(sample).
I suspect bisecting k-means is notable enough and different enough for
inclu
Andreas,
There isn't necessarily a linkage function defined, at least in the sense
of agglomerative clustering, since this is not comparing clusters to merge
but rather breaking them up. The clusters are split using another
clustering algorithm supplied by the caller. The most common one that I've
In my experience it is not very helpful to talk about agglomerative vs
divisive algorithms, as that is often more of an implementation detail.
Single-link agglomerative clustering for example is often implemented by
computing the spanning tree and then cutting it.
So it is more of a question wha
