Graph-based boosting algorithm to learn labeled and unlabeled data. (October 2020)

• Record Type: Journal Article
• Authors: Liu, Zheng; Jin, Wei; Mu, Ying
• Is Part Of: Pattern recognition. Volume 106(2020:Oct.)
• Publication Date: 2020-10
• Subjects: Graph -- Boosting -- Semi-supervised learning -- Imbalance learning; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2020.107417
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: Some ensemble learning algorithms were proposed to exploit the information of unlabeled data. These methods had to learn the samples with pseudo-labels due to the scarcity of labeled data. But it is inevitable for the samples with pseudo-labels to bring wrong information during training process. In this paper, we propose a novel graph-based boosting algorithm (GBB) to learn labeled and unlabeled data. And pseudo-labels will not occur during training process. GBB is a framework combining many models linearly and the similarity matrix of samples is transformed during training process. We also extend GBB, termed as weighted GBB (WGBB), to learn imbalanced data. Experimental results illustrate that GBB can achieve a competitive performance and WGBB has an obvious advantage to handle classification problem of imbalanced data, comparing with other related algorithms. Abstract: Ensemble learning is an effective technique to learn the information of data by combining multiple models. But usually the combined models are supervised learning algorithms which need a lot of labeled data to tune their parameters. Some ensemble learning algorithms were proposed to exploit the information of unlabeled data. These methods had to learn the samples with pseudo-labels due to the scarcity of labeled data. But it's inevitable for the samples with pseudo-labels to bring wrong information during training process. In this paper, we will propose a novel graph-based boosting (GBB) algorithm to learn labeled and unlabeled data. GBB is a framework combining many models linearly. And pseudo-labels will not occur during training process. GBB will assign a new weighting vector for the labeled samples and a transformed similarity matrix for all samples to train the combined model at each iteration. We also extend GBB, termed as weighted GBB (WGBB), to learn imbalanced data by adding a weighting vector for the labeled data. Finally, 14 relatively balanced datasets and 22 imbalanced datasets are used to validate the performances of GBB and WGBB respectively. Experimental results illustrate that GBB can achieve a competitive performance and WGBB has an obvious advantage to handle classification problem of imbalanced data, comparing with other related algorithms.