Transfer learning with adaptive regularizers

Abstract

The success of regularized risk minimization approaches to classification with linear models depends crucially on the selection of a regularization term that matches with the learning task at hand. If the necessary domain expertise is rare or hard to formalize, it may be difficult to find a good regularizer. On the other hand, if plenty of related or similar data is available, it is a natural approach to adjust the regularizer for the new learning problem based on the characteristics of the related data. In this paper, we study the problem of obtaining good parameter values for a ℓ2-style regularizer with feature weights. We analytically investigate a moment-based method to obtain good values and give uniform convergence bounds for the prediction error on the target learning task. An empirical study shows that the approach can improve predictive accuracy considerably in the application domain of text classification. © 2011 Springer-Verlag.