Online/Batch generalized linear models under square loss

Online (competitive)/batch prediction using generalized linear models under square loss

This is the package for online (competitive)/batch prediction using generalized linear models under square loss. The algorithms are derived using the Aggregating Algorithm.

The algorithms have guarantees on the cumulative square loss for the worst case when applied in online fashion in comparison with the best model from the class [1].

The variable regressed should lie in [0,1], thus the program is a tool for two-class classification or for bounded regression.
Four possibilities are provided: linear regression, logistic regression, probit regression, comlog regression. Other functions can be easily added/used.

The models are developed and first applied in [1], the competitor to linear regressor (AAR) was first suggested in [2].

FIle examplepredict.m contains an example of use. The data set is the wine data set available from UCI Machine Learning Repository.
Two first clases are taken, vectors are randomly permuted, features are normalized to have zero mean and maximum absolute value 1.
This particular problem is not very suitable for online regression, so the data set just illustrates how to use the program.

References:
1. Fedor Zhdanov and Vladimir Vovk. Competitive online generalized linear regression under square loss, to appear in ECML 2010 proceedings.
2. Vladimir Vovk. Competitive on-line statistics. International Statistical Review, 69:213–248, 2001.

(C) Fedor Zhdanov, 2010.