`predict_coordinates` suggestions about q2-sample-classifier HOT 7 CLOSED

would it be better to do this as a multilabel regression?
currently each dimension is predicted separately

@BenKaehler , could you offer some advice on this? I think multilabel/multioutput regression could theoretically increase accuracy if covariation between each target is taken into account. As far as I can tell from the sklearn docs, MultiOutputRegressor does not do this, however, so we may be stuck.

from q2-sample-classifier.

Hi Everyone, my understanding is that scikit-learn doesn't offer anything more sophisticated than providing multiple independent regressors for multioutput problems.

An alternative (that I didn't get time to look into for the feature-classifier) is to use structured learning. Apparently there are several possible libraries to use for structured learning, but I though the best-looking Python one was pystruct. It does have some licensing issues though.

Let me know if you would like me to look into it any further, or if you find a better solution!

from q2-sample-classifier.

It does have some licensing issues though

From pystruct's github page, it appears to be BSD-2. Would this be incompatible with BSD-3?

from q2-sample-classifier.

It's not pystruct that's the problem, it's the libraries that it relies upon. See pystruct/pystruct#192 for details.

from q2-sample-classifier.

Have changed latitude and longitude to axis1_category and axis2_category.

Utilizing pystruct is still unresolved... will leave this issue open while I explore options for structured learning.

from q2-sample-classifier.

it appears to be BSD-2. Would this be incompatible with BSD-3?

No, it is compatible.

from q2-sample-classifier.

predict_coordinates is being removed from this plugin, so I am deleting this issue.

from q2-sample-classifier.