On Wed, Oct 19, 2011 at 05:32, Lars Buitinck <[email protected]> wrote:
> Hi everyone,
>
> I was wondering what the LogisticRegression class in
> sklearn.linear_modules implements. Its docstring says just "logistic
> regression", but I stumbled upon the liblinear authors' paper on LR
> [1], and they claim to have developed a fast multiclass LR/MaxEnt
> training algorithm as well. Does our implementation do LR or ME?

Unregularized logistic regression and maximum entropy are duals of
each other (which means they are the same thing), and in fact all
implementations of maximum entropy optimize in the dual (because the
primal has one variable per training point, one constraint per
feature, and no way of generalizing to test data). Regularization in
logistic regression changes the duality a bit, however.

Maxent is usually older terminology that used to be favored in the NLP
community but has recently been phased out due to confusion and a
mismatch with the way people actually implement things.

These days you will see people optimizing the log loss, which for
binary classification is log(1 + exp(-y W dot X)) and for multiclass
classification is W_y dot X - log(sum_y' W_y' dot X).
-- 
 - Alexandre

------------------------------------------------------------------------------
All the data continuously generated in your IT infrastructure contains a
definitive record of customers, application performance, security
threats, fraudulent activity and more. Splunk takes this data and makes
sense of it. Business sense. IT sense. Common sense.
http://p.sf.net/sfu/splunk-d2d-oct
_______________________________________________
Scikit-learn-general mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/scikit-learn-general

Reply via email to