Regularization on 3-walks in associative graph for unsupervised domain adaptation

Abstract

We propose a new domain adaptation method that adjusts feature space itself. In order to adjust the distorted space by the classifier, the relation between the data feature and the class is defined and designed to optimize it. We define a 3-partite graph composed of source domain data, test domain data, and class. The connection between them is defined as a transition probability function between each group. We use a regularization loss for the graph to optimize the connection between classes only self-looped. This is an end-to-end unsupervised learning method that can be easily attached to various domain adaptation methods and is computationally efficient. The results of the domain adaptation experiment between several digit datasets show that the proposed method shows good performance and is a way to change the decision boundary of the classifier because the two domains are separated in the embedding space. In addition, this approach demonstrates improved test performance for the source domain, reinforcing the generalization of the source domain.