search

alashworth / test-issue-import

0 stars 0 forks source link

Add multinomial_logit #97

Open alashworth opened 5 years ago

alashworth commented 5 years ago

Issue by demodw Thursday May 26, 2016 at 15:50 GMT Originally opened as https://github.com/stan-dev/stan/issues/1897

Summary:

Add a function that matches the existing bernoulli_logit and categorical_logit.

Description:

This would add a convenience function usable in multi-outcome logistic regression problems. For models with larger data sets it would add a significant speedup, given that the current multinomial function has to be called every single iteration, since it currently only accepts a single vector.

Current behavior

Assume there are N data points and K different outcomes, y is a NxK matrix of integers, beta is a K vector with some combination of covariates and predictors for each outcome,

for (n in 1:N) {
    ...
    y[n] ~ multinomial(softmax(beta))
}

Wanted behavior

y ~ multinomial_logit(beta)

In this case beta would be a NxK matrix with some combination of covariates and predictors for each outcome.

Additional Information:

I could not find any open issues for this missing feature. Also, I am not sure if this should be added as an issue to the stan-dev/math repository.

Current Version:

v2.9.0

alashworth commented 5 years ago

Comment by bob-carpenter Thursday May 26, 2016 at 18:56 GMT

It will need to be implemented in stan-dev/math, at which point adding it to Stan is a a couple lines of code and some doc.

There's a categorical_logit which can be followed here.

The big speedup will be when there's a single beta vector and an array of y outcomes; then the softmax will only need to be computed once and it and the derivates can be reused. If beta and y are both arrays, it won't be much faster than a loop, because there's no underlying floating-point calculations to share.