A Cross-Ciew Re-Alignment Approach for Incomplete Multi-View Contrastive Clustering

カテゴリ: 部門大会

論文No: OS4-1-4

グループ名: 【C】2024年電気学会電子・情報・システム部門大会

発行日: 2024/08/28

タイトル(英語): A Cross-Ciew Re-Alignment Approach for Incomplete Multi-View Contrastive Clustering

著者名: Ding Yu（福井大学）,Hotta Katsuya（岩手大学）,Gu Chunzhi（豊橋技術科学大学）,Yu Jun（新潟大学）,Zhang Chao（富山大学）

著者名(英語): Yu Ding (University of Fukui),Katsuya Hotta (Iwate University),Chunzhi Gu (Toyohashi University of Technology),Jun Yu (Niigata University),Chao Zhang (University of Toyama)

要約(日本語): The task of incomplete multi-view clustering aims at learning a common representation from multiple views including missing data, while dividing into different clusters. Existing approaches only consider the missing data and rely heavily on the property of view alignment, whereas the multi-view data would be only partially aligned. Thus, we propose a novel incomplete multi-view clustering framework that learns view consistency from partially aligned data to further tackle the view-unaligned problem. Our key insight is to learn view-invariant features by redividing the unaligned sample pairs into true-negative and false-negative pairs. We then realign the unaligned part via our trained model to facilitate the clustering task. Extensive experiments show the effectiveness of our model.