An Adversarial Approach to Importance Weighting for Domain Adaptation

Abstract

Most supervised learning methods assume that the training and test data points are generated from the same distribution. Therefore, they face major challenges when this assumption does not hold. Domain adaptation techniques aim to address this issue by adapting the learned models to new distributions. Covariate shift is a common assumption in domain adaptation, where the training and test distributions only differ in the marginal distributions. A common idea to tackle covariate shift is estimating the importance weights of the training data points using unlabeled data from the source and target distributions and then training the classifier using importance-weighted risk minimization. Existing methods for estimating the importance weights are kernel-based which scale poorly with dataset size and underperform on high-dimensional data. This work proposes a novel method for estimating the importance weights using generative adversarial networks. There are two neural networks used in this framework which we call the weighting and the discriminator networks. These networks are jointly trained using an adversarial learning scheme. We designed a benchmark for assessing classification performance under various forms of distribution shift and evaluated our method in this framework. We observe that while our method effectively estimates the importance weights, the improvements we get in the domain adaptation task depend on the discrepancy of the domains.