Hey Stefan.
I would expect that to depend on the prior.
It could either be a bug or an issue with the variational inference.
Maybe comparing against an MCMC implementation might be helpful?
Though if that works, I'm not sure what the conclusion would be tbh.
(I hate debugging variational inference, I can't get the hang of it)
Can you check the estimated covariance? what is it?
The samples that you're showing are from all 100 components, right?
Cheers,
Andy
On 2/6/19 1:34 PM, Stefan Ulbrich via scikit-learn wrote:
Hello,
I think I might have found a bug in the BayesianGaussianMixture–or at
least encountered a behavior that I was not expecting. The problem
occurs when having clusters with small extent (in my case, it is 2D
geographic data) that are far away from each other. While the means
and their number are determined correctly, the co-variance matrices
are not (at least compared to the regular GMM): They are are much
wider and point towards the mean of the cluster centers.
A minimal example and visualization can be seen on a stackoverflow
question I opened.
https://stackoverflow.com/q/54524283
So my question is whether the results of GMM and BGMM should be
similar or this is the expected behavior (and why)?
Thanks in advance for an answer and best wishes
Stefan
_______________________________________________
scikit-learn mailing list
scikit-learn@python.org
https://mail.python.org/mailman/listinfo/scikit-learn
_______________________________________________
scikit-learn mailing list
scikit-learn@python.org
https://mail.python.org/mailman/listinfo/scikit-learn
