Saturday, April 29, 2006

Naive Bayes vs. Logistic Regression

There's often confusion as to the nature of the differences between Logistic Regression and Naive Bayes Classifier. One way to look at it is that Logistic Regression and NBC consider the same hypothesis space, but use different loss functions, which leads to different models for some datasets.

To see that both logistic regression and naive bayes classifier consider the same hypothesis space we can rewrite Naive Bayes density as follows (restricting attention to binary domain):

where

Alternatively you can get exponential family parametrization from the observation that Naive Bayes model has no unshielded colliders, so undirected model obtained by dropping the arrows is equivalent on the set of positive densities.

From here you can rewrite it as product of conditional and marginal distributions

You can see that the conditional term is equivalent to logistic regression, so every possible conditional density that can be modelled by logistic regression can be modelled by Naive Bayes by setting \phi's aribrarily.

So both logistic regression and Naive Bayes have the same hypothesis space, but optimize different objective functions. In particular logistic regression maximizes

whereas Naive Bayes maximizies

You can see that the second term involves both phis and thetas, so conditional and marginal likelihoods are coupled. If empirical density is realizable by our model, then this coupling doesn't matter -- both conditional and marginal terms can achieve their respective maxima so Naive Bayes and Logistic Regression will produce the same estimate. However that not necessarily true when empirical density is unrealizable -- you may have to compromise between achieving high conditional likelihood and high marginal likelihood.

To get a generatively trained model that is identical to conditionally trained Naive Bayes, Minka suggests treating parameters in conditional and joint as separate sets.

So now we can introduce n new parameters so that conditional and marginal densities become decoupled.

It's interesting to see what is the new marginal density of our model. The marginal density marginalizes out Y. Since Y is the only separator in original graph, the resulting graph will be fully connected.

So if we were to restrict attention to linear exponential families, we'd have to introduce 2^n features. I'm wondering if it's possible to infer that fact by looking at the form of the density. You'd have to show that the size of sufficient statistic is at least 2^n, or alternatively show that the smallest linear space that embeds the unnormalized log densities has dimension 2^n

If x_i are real or complex valued, one could show that the smallest linear space that embeds the set above has to be infinite dimensional, by noting that since log(1+exp(ax)) has singularities at i pi a, so one can always construct linearly independent elements by choosing a accordingly. I wonder, are there similar tricks that would work for discrete x?

(derivations)

Anonymous said...

Hi Yaroslav,

I'm very interested in the post you published under the link below.

http://yaroslavvb.blogspot.com/2006/04/naive-bayes-vs-logistic-regression.html

(naive bayes vs. regression)

However, the link doesn't show the formulas.
Can you please send me the version with the formulas?

Shira
shirrra@gmail.com

IN-omaly said...

amazing. Thanks.

Really helping with my ML course.

daryoush said...

When you say:

"...both logistic regression and naive bayes classifier consider the same hypothesis space"

I am not sure I understand the significance of "same hypothesis space" comment. When (or which algorithms) would have different hypothesis space.

Thanks,

Daryoush

Arthur Mendoza said...

Actually there has some similarities between these two but none of them are not that bad than others in quality maters. more that I love for the helpful information and its really amazing.

Charles Kee said...

The post admission of jamb is here now and there has form for the admission and you should fill it up before reading the rules. https://pathology.residencypersonalstatements.net/pathology-residency-personal-statement-sample/ that is the best one place to know about the quality writing rules.

sina said...

It is one of the easiest things to do the cover letter as the electronics sector is not that tough to functionized here. website go here and you'll be get and helpful ideas on the subject of academic papers writing.

Geetha Devi said...

I think things like this are really interesting. I absolutely love to find unique places like this. It really looks super creepy though!!

Best Machine Learning institute in Chennai | machine learning with python course in chennai | best training institute for machine learning

smak said...

We can understand the formulas with the help of https://www.lorservice.com/our-recommendation-letter-service/letter-of-recommendation-for-nursing-school-writing-service/ because these formulas are for professionals and the professionals know how to use those and where to use those.

Jane said...

There is not just confusion but the difference is not clear that creates confusion but now https://www.nursingpaper.com/our-services/nursing-reflective-journal/ is there to clear such confusion and that can help the teachers to learn what is required in reality.