A bi-level metric learning framework via self-paced learning weighting. (July 2023)

• Record Type: Journal Article
• Authors: Yan, Jing; Wei, Wei; Guo, Xinyao; Dang, Chuangyin; Liang, Jiye
• Is Part Of: Pattern recognition. Volume 139(2023)
• Publication Date: 2023-07
• Subjects: Metric learning -- Self-paced learning -- Adaptive neighborhood -- Weighting tuples; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2023.109446
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: We propose a novel bi-level framework for learning an effective Mahalanobis distance metric. We introduce an implementation of the proposed framework based on the self-paced learning method and design the corresponding optimization algorithm. The robustness of the proposed model is improved by reducing the effect of noisy samples on the model via a self-paced learning regular term. Experimental results demonstrate that the proposed algorithm outperforms the state-of-the-art methods on several benchmark datasets. Abstract: Distance metric learning (DML) has achieved great success in many real-world applications. However, most existing DML models characterize the quality of tuples on the tuple level while ignoring the anchor level. Therefore, the models are less accurate to portray the quality of tuples and tend to be over-fitting when anchors are noisy samples. In this paper, we devise a bi-level metric learning framework (BMLF), which characterizes the quality of tuples more finely on both levels, enhancing the generalization performance of the DML model. Furthermore, we present an implementation of BMLF based on a self-paced learning regular term and design the corresponding optimization algorithm. By weighing tuples on the anchor level and training the model using tuples with higher weights preferentially, the side effect of low-quality noisy samples will be alleviated. We empirically demonstrate that the effectiveness and robustness of the proposed method outperform the state-of-the-art methods on several benchmark datasets.