In a variational Logistic regression approach, we approximate a logistic sigmoid function to a Gaussian Distribution by means of its lower bound.

It can be solved using an EM algorithm, given that we work with a fixed prior and known initial parameters (which we usually do not know).

My question is: What is the merit of this approach? since common logistic regression has a closed solution and is faster to implement than an EM approach.

What can we learn from the data by using the variational approach rather than the straight forward approach.

Thanks

asked May 30 '11 at 02:34

Leon%20Palafox's gravatar image

Leon Palafox ♦
40857194128

Can you give some source where this method is described?

(May 30 '11 at 05:16) Andreas Mueller

Sure Thing, it's in Bishop's chapter on Variational Inference (10) sub-chapter 6 , variational logistic regression (10.6)

(May 30 '11 at 05:37) Leon Palafox ♦

Are you sure that common logistic regression has a closed form solution?

(Jun 01 '11 at 09:44) Denzel

Logistic regression does not have a closed form solution.

(Jun 01 '11 at 10:32) Ian Goodfellow

Indeed, it was my mistake, still, MAP Logistic regression is way simpler than variational logistic regression

(Jun 02 '11 at 01:11) Leon Palafox ♦

2 Answers:

The variational approach fits a Bayesian logistic regression model. This means that you do not only get the maximum likelihood solution for the weights, you get a distribution over weights. And even more: you can integrate out the weights and make predictions considering "every possible weight-vector". If you think this is a good thing probably depends on your opinion on Bayesian methods. In practice, this amounts to "automatic regularization" - where you hope that you can either integrate out hyper parameters, hope to have some idea what you should use, or use non-informative priors.

answered May 30 '11 at 05:49

Andreas%20Mueller's gravatar image

Andreas Mueller
2686185893

Adding to Andreas Muller's answer, the main advantage of variational logistic regression for me is that it is easier to adapt to more complex models as a black box. So you can, for example, do variational inference on a supervised LDA model with logistic regression using the topics as features, and it will all fall out nicely, while trating the logistic function analytically would be a fair amount of work.

answered May 30 '11 at 12:51

Alexandre%20Passos's gravatar image

Alexandre Passos ♦
2554154278421

You mean since a variational approach allows us to model a sigmoid as a normal, and thus we can use Conjugate properties?

(May 31 '11 at 01:57) Leon Palafox ♦
Your answer
toggle preview

powered by OSQA

User submitted content is under Creative Commons: Attribution - Share Alike; Other things copyright (C) 2010, MetaOptimize LLC.