Joint distribution matching embedding for unsupervised domain adaptation

摘要

When the distributions between the source (training) and target (test) datasets are different, the performance of classical statistical learning methods degrades significantly. Domain adaptation (DA) aims at correcting this distribution mismatch and narrowing down the distribution discrepancy. Existing methods mostly focus on correcting the mismatch between the marginal distributions and/or the class-conditional distributions. In this paper, we assume that the distribution mismatch in domain adaptation is the joint distribution mismatch, and propose an Extended Maximum Mean Discrepancy (EMMD) metric to measure the distance between joint distributions. Based on this metric, we propose the Joint Distribution Matching Embedding (JDME) approach, which finds a mapping matrix to project the samples into a latent space, where the EMMD between the source and target joint distributions is minimized. The resultant orthogonal-constrained optimization problem can be solved in the form of an unconstrained problem on the Grassmann manifold. After the joint distribution matching, we can expect the classical statistical learning methods to perform well on the target dataset. Experiments on object recognition, face recognition, and spam filtering demonstrate that our method statistically outperforms the state-of-the-art shallow methods and is also on par with the deep learning methods. (c) 2020 Elsevier B.V. All rights reserved.