Auto-weighted sample-level fusion with anchors for incomplete multi-view clustering. (October 2022)

• Record Type: Journal Article
• Authors: Yu, Xiao; Liu, Hui; Lin, Yuxiu; Wu, Yan; Zhang, Caiming
• Is Part Of: Pattern recognition. Volume 130(2022)
• Publication Date: 2022-10
• Subjects: Incomplete data -- Multi-view clustering -- Anchor -- Auto-weighted -- Large-scale; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2022.108772
• Languages: English
• ISSNs: 0031-3203

Abstract

Abstract : highlights: An anchor graph based method is proposed for incomplete multi-view clustering. It can be applied to large-scale multi-view datasets. The instance-to-anchor similarity fusion is implemented at the sample level in an auto-weighted way to get the optimal weight for the corresponding view of each sample. Experiments on 32 datasets with eight baselines show the superiority of our method. Even on datasets with only 10% complete samples, ASA-IC still performs well. Graphical abstract: Abstract: Aiming at solving the problem of clustering in the multi-view datasets which include samples with information missing in one or more views, incomplete multi-view clustering has received considerable attention. However, most studies can not get satisfying accuracy and efficiency when dealing with datasets in which a considerable number of instances are missing in partial views. To address this problem, a method named Auto-weighted Sample-level Fusion with Anchors for Incomplete Multi-view Clustering (ASA-IC) is proposed in this paper. It designs an auto-weighted sample-level fusion strategy, which realizes the optimized conversion from the individual instance-to-anchor similarity learning to the concensus instance-to-anchor similarity matrix construction. ASA-IC can not only handle incomplete samples and effectively explore the relationship between each instance and anchors, but also deal with various incomplete clustering situations and be applied in large-scale datasets as well. Besides, experiments on 5 complete datasets and 27 incomplete ones illustrate its effectiveness quantitatively and qualitatively.