Predict top-k sequences

[jbingel]

Thanks for a fine piece of software! It would be super neat if one were able to have the estimators predict the top k tag sequences for some input, in other words to retrieve several candidates.

I'm thinking something along the lines of scikit-learn's predict_proba() function. Does the Viterbi implementation used in pystruct allow for this?

[amueller]

Hi Joachim.
Unfortunately, that is very non-trivial. There is active research on
that, but I don't have the current bandwidth to work on that.
If you implement something like predict_proba, the shape of the
output array would be exponential, as there are exponentially many
possible labellings.
[Also, it would be more like "decision function" because pystruct
doesn't produce normalized probabilities].
If you look into top-k predictions for structured SVMs, you'll be able
to find some publications, I think.

Cheers,
Andy

On 02/27/2016 04:55 PM, Joachim Bingel wrote:

Thanks for a fine piece of software! It would be super neat if one
were able to have the estimators predict the top /k/ tag sequences for
some input, in other words to retrieve several candidates.

I'm thinking something along the lines of scikit-learn's
|predict_proba()| function. Does the Viterbi implementation used in
pystruct allow for this?

—
Reply to this email directly or view it on GitHub
#175.

[jbingel]

Thanks for the reply. Yeah I was assuming that, but thought I'd give it a shot. :) Thanks also for the pointer to structured SVMs.