Ensemble selection with joint spectral clustering and structural sparsity. (November 2021)

• Record Type: Journal Article
• Authors: Wang, Zhenlei; Zhao, Suyun; Li, Zheng; Chen, Hong; Li, Cuiping; Shen, Yufeng
• Is Part Of: Pattern recognition. Volume 119(2021)
• Publication Date: 2021-11
• Subjects: Ensemble selection -- Structural sparsity -- Unsupervised selection -- Spectral clustering -- Dynamic and static -- Robustness; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2021.108061
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: First unsupervised ensemble selection method with joint clustering and sparsity. A prediction space leveraging the prediction power of an ensemble is the basis of the method. The data in the prediction space, which globally describe the prediction distribution of the ensemble, are unlabeled. Method's ensemble results are robust to test instances and uses less space. Abstract: Generally, ensemble selection techniques are split into two categories: dynamic and static. Static ensemble selection selects a fixed subset of the original ensemble which improves the space complexity but is not flexible to each test instance. Dynamic ensemble selection selects base learners on-the-fly according to each test instance but it does not significantly improve the complexity. Currently, there is no ensemble selection technique that is robust to the test instances as well as improves space complexity. To narrow this gap, we propose a novel static ensemble selection method, called Ensemble Selection with Joint Spectral Clustering and Structural Sparsity. This method integrates spectral clustering and structural sparsity into a joint framework whose ensemble selection result is robust to test instances and consumes less space. Using 25 datasets from KEEL and UCI, we demonstrate the effectiveness of our proposed algorithm and its promising performance compared to that of other state-of-the-art algorithms.