ENH: add y-intercept for logistic regression

[facaiy]

Hi, the PR is aimed at adding intercept for logistic regression #142 .

The work is not done yet (mostly finished in fact), but feedback is necessary for me to determine whether to go further or not. Hence, the PR is opened.

Mathematical Derivation

Given:

• linear model: f(x) = w^T x + b
• loss function: L(y, f(x))

the derivative of b is: \frac{dL}{df}

Implementation

The PR is inspired by:

• spark: LogisticRegression.scala
• sklearn: logistic.py

How to test

• code review by expert.
• Perhaps some tests are needed, however, more feedback is needed for me now.

[TAAAN]

Thanks for your PR.
When preprocessing train data, set the first dim 1 for every instance, then the first dim of w is b.

[andyyehoo]

@facaiy it's a good idea. Just go on with this PR and finish it. We will review and merge it with certain modification.

[facaiy]

OK. I'll ping you when finished.

[facaiy]

The PR is tested on a4a dataset.

There are two questions:

1. The package failed to submit its tasks to YARN. I believe the bug is introduced from the latest branch-1.0.0, instead of the PR. I will create a new issue to report it later.

2. I picked those commits to commit 3ff4329825fd355118384d5be438dd132f1015c3, and everything is OK. However, the result of experiment on a4a dataset implies that AUC decreased when y-intercept is used.

use y-intercept	iteration	auc	duration	reg_l2	learn rate
No	10	0.8991570991465441	47s	0	4
Yes	10	0.898142663326485	50s	0	4

[facaiy]

Many thanks, @andyyehoo @TAAAN .