COMB (combtrain/combclassify) is a libary for learning L1 regularized logistic regression model with combination features.

Main features are

Automatic extraction of effective combination features Compact model (due to L1 regularization) Fast inference Specialized for large-scale, but sparse learning task, such as natural language processing.

When an original problem cannot be separated linearly, combination features (products of several original features) are very effective. In many cases, these features are added by manually. Althoguh kernel trick can capture such non-linear phonomena implicitly, the inference is very heavy, and the model is not interpretable.

COMB supports automatic extraction of effective combination features by using column-generation algorithm, efficient filtering techniques. Currently, COMB supports the libsvm-style data representation.