Abstract
This paper focuses on the on-line learning method called Adaptive Regularization of Weight Vectors (AROW). AROW has two advantages compared to other on-line learning methods. First, it is robust to label noise. Second, we can obtain stable models by taking confidence intervals of parameters into consideration. However, AROW cannot perform feature selection. This paper proposes a novel method by combining AROW with lasso. We also employ the coordinate descent algorithm to estimate parameters, which enables us to speed up our algorithm. We confirm the effectiveness of our proposed method by some numerical experiments.
