Possibilistic Distribution Distance Measure: Robust Domain Adaptation Learning Method

Abstract

Domain adaptation (DA) aims to solve the problem of inconsistent distribution between training dataset and test dataset, which has attracted extensive attention. Most of the existing DA methods solve this problem by the maximum mean discrepancy (MMD) criterion or its variants. However, the noise data may lead to a significant drift of domain mean, which will reduce the performance of MMD and its variants to some extent. To this end, this paper proposes a robust domain adaptation method with possibilistic distribution distance measure. Firstly, the traditional MMD criterion is transformed into a new possibilistic clustering model, which aims to reduce the impact from noise data. This paper constructs a robust possibilistic distribution distance measure (P-DDM) criterion. It further improves the robust effectiveness of domain distribution alignment by adding the fuzzy entropy regularization term. Secondly, a domain adaptation visual classifier based on P-DDM (C-PDDM) is proposed. It adopts a graphical Laplacian matrix for preserving the geometric consistency of data in source domain and target domain. It can improve the label propagation performance. In order to improve generalization, it maximizes the use of source domain discrimination information to minimize the domain discrimination error. Theoretical analysis confirms that the proposed P-DDM is an upper bound of the traditional distribution distance measurement method MMD criterion under certain conditions. Therefore, minimizing the P-DDM can effectively optimize the MMD objective. Finally,  it is compared with several representative domain adaptation methods, and the experimental results  on 6 visual benchmark datasets (Office31, Office-Caltech, Office-Home, PIE, MNIST-UPS, and COIL20) show that the proposed method achieves an average improvement of about 5% on generalization performance and an average improvement of about 10% on robustness performance.

Keywords

• domain adaptation (da); probabilistic clustering; maximum mean discrepancy (mmd); fuzzy entropy