Domain adaptation via incremental confidence samples into classification. Issue 1 (24th August 2021)

• Record Type: Journal Article
• Authors: Teng, Shaohua; Zheng, Zefeng; Wu, Naiqi; Fei, Lunke; Zhang, Wei
• Is Part Of: International journal of intelligent systems. Volume 37:Issue 1(2022)
• Page Start: 365
• Page End: 385
• Publication Date: 2021-08-24
• Subjects: adaptive adjustment -- domain adaptation -- joint probability distribution -- selective pseudo label -- transfer learning; Artificial intelligence -- Periodicals; Expert systems (Computer science) -- Periodicals; Intelligence artificielle -- Périodiques; Systèmes experts (Informatique) -- Périodiques; 006.3
• DOI: 10.1002/int.22629
• Languages: English
• ISSNs: 0884-8173

Abstract

Abstract: To accurately recognize similar objects in different domains, the key for domain adaptation is to learn new metrics so as to minimize the discrepancy of two domains. Recent works utilize joint probability domain adaptation to tackle this problem but get poor performance due to poor discriminability or transferability of data sets. The inaccurate pseudo‐labeling in the feature subspace can lead to a chain reaction of errors during iterations, and varieties of the joint probability distribution values further aggravate the miscalculation. To cope with the above problems, this paper proposes a unified framework by introducing Incremental Confidence Samples into Classification (ICSC). ICSC includes both incrementally labeling and adaptively adjusting. With the increase of confidence samples in each iteration, incrementally labeling is used to reduce error accumulations and progressively guarantee good classification performance effectively. Moreover, the tradeoff weight between within‐class and between‐class distance is adaptively adjusted according to the importance of transferability and discriminability. Consequently, the discrepancy minimization of within‐class and the discrepancy maximization of between‐class are achieved. Extensive experiments on several benchmark data sets demonstrate the effectiveness of the proposed method over the state‐of‐the‐art methods.