Similarity-based domain adaptation network

Abstract

Domain adaptation utilizes labeled source domains to solve classification problems in the unlabeled target domain. Previous domain adaptation methods consider global domain adaptation while neglecting class-wise information, thus leading to poor transfer performance. In recent years, many researchers studied class-wise domain adaptation, the focus of which is to accurately align the distribution of different domains. However, these methods cannot distinguish samples with different similarities during domain adaptation, which results in the inaccurate matching of different domains. Therefore, this paper proposes a Similarity-Based Adaptation Network (SBAN), which optimizes Similarity-Based Domain Discrepancy (SBDD) that models similarity-based intra-domain and inter-domain discrepancies, and proposes an alternating update strategy to train the SBAN. Specifically, we assign different weights to samples with different similarities, and minimize the similarity-based inter-domain discrepancy to make similar samples dominate the distribution alignment across domains while alleviating the effect of dissimilar samples, and minimize the similarity-based intra-domain discrepancy to align dissimilar samples with similar samples within the same domain to learn more discriminative representations. Extensive experiments on four widely used benchmark datasets show that SBAN performs better than several latest domain adaptation methods.