Right now, this package includes Online Gradient Descent Logistic Regression (OGDLR) and Follow the (proximal) Regularized Leader (FTRLprox) algorithms.

Some features:

• The methods work with extremely sparse data (all treated categorically) by using dictionary for storage or hashing trick. This allows to train very sparse feature sets without exhausting memory.
• The interface is similar to that of scikit learn.
• Cross-validation is made on the fly.

All this is work in progress, use at your own risk.

For paper, see (here)[http://www.eecs.tufts.edu/~dsculley/papers/ad-click-prediction.pdf].