An empirical study about online learning with generalized passive-aggressive approaches

This work aims at exploring different approaches to online learning that can be grouped under the common denomination of passiveaggressive techniques. In particular, we comparatively explore the original passive-aggressive formulation along with a recently proposed leastsquares extension and a new proposal in which aggregate updates corresponding to small groups of labelled examples are considered instead of single samples. Preliminary results show that extended algorithms perform at least as good as basic ones in the long term but exhibit a smoother and more robust behavior along learning iterations.