Should mixture models have a clusterer-compatible interface

[jnothman]

Mixture models are currently a bit different. They are basically clusterers, except they are probabilistic, and are applied to inductive problems unlike many clusterers. But they are unlike clusterers in API:

• they have an n_components parameter, with identical purpose to n_clusters
• they do not store the labels_ of the training data
• they do not have a fit_predict method

And they are almost entirely documented separately.

Should we make the MMs more like clusterers?

[spinicist]

In my opinion, yes.

I wanted to compare K-Means, GMM and HDBSCAN and was very disappointed that GMM does not have a fit_predict method. The HDBSCAN examples use fit_predict, so I was expecting GMM to have the same interface.

[amueller]

I think we should add fit_predict at least. I wouldn't rename n_components.

[Eight1911]

I would like to work on this!

[amueller]

@Eight1911 go for it. It is probably relatively simple but maybe not entirely trivial.

[g-walsh]

@Eight1911 Mind if I take a look at this?

[haoranShu]

@Eight1911 Do you mind if I jump in as well?