Generalization Bounds Derived IPM-Based Regularization for Domain Adaptation

Abstract EN

Domain adaptation has received much attention as a majorform of transfer learning.

One issue that should be considered indomain adaptation is the gap between source domain andtarget domain.

In order to improve the generalization abilityof domain adaption methods, we proposed a frameworkfor domain adaptation combining source and target data,with a new regularizer which takes generalization boundsinto account.

This regularization term considers integralprobability metric (IPM) as the distance between thesource domain and the target domain and thus can boundup the testing error of an existing predictor from theformula.

Since the computation of IPM only involvestwo distributions, this generalization term is independentwith specific classifiers.

With popular learning models,the empirical risk minimization is expressed as a generalconvex optimization problem and thus can be solved effectivelyby existing tools.

Empirical studies on synthetic data forregression and real-world data for classification show theeffectiveness of this method.