Probabilistic similarity preservation for distribution discrepancy reduction in domain adaptation

Abstract

Many existing domain adaptation methods attempt to reduce inter-domain differences without accounting for the varying weights of the feature vectors in the two domains. Since the labels estimated for the target domain data may not be accurate when calculating the conditional probability distributions, the domain adaptation methods based on these distributions are negatively impacted by these noisy labels. In this paper, a new domain adaptation method is proposed to reduce inter-domain discrepancy by considering the group relationships of samples based on the potential similarities derived from the distances between samples in source and target domains. To achieve this goal, a new similarity matrix is introduced to extract prototypes from the target domain represented as a combination of samples feature vector. These prototypes are extracted by trying to preserve the probabilistic similarities to the source domain labeled samples, limiting the impact of noisy labels when calculating the conditional probability distributions in the target domain. Following this, the conditional probability distribution divergences between the two domains for each class are reduced by imposing a common distribution function on these feature vectors. Extracting the prototypes from the target domain as well as considering similar distribution form for the feature vectors of the two domains, allow the discrepancy between the two domains to be reduced more effectively. Three different datasets are used to evaluate the proposed method: Office-31, Office-Home, and ImageCLEF. Experimental results on these datasets demonstrate that our approach achieves considerably higher accuracy compared to other recent related works. © 2025 Elsevier Ltd