[1]ZHANG Zhihui,YANG Yan,ZHANG Yiling.Deep mutual information maximization method for incomplete multi-view clustering[J].CAAI Transactions on Intelligent Systems,2023,18(1):12-22.[doi:10.11992/tis.202203051]
Copy
CAAI Transactions on Intelligent Systems[ISSN 1673-4785/CN 23-1538/TP] Volume:
18
Number of periods:
2023 1
Page number:
12-22
Column:
学术论文—机器学习
Public date:
2023-01-05
- Title:
-
Deep mutual information maximization method for incomplete multi-view clustering
- Author(s):
-
ZHANG Zhihui; YANG Yan; ZHANG Yiling
-
School of Computing and Artificial Intelligence, Southwest Jiaotong University, Chengdu 611756, China
-
- Keywords:
-
data mining; clustering; incomplete multi-view clustering; multi-view representation learning; deep learning; autoencoder; mutual information; self-paced learning
- CLC:
-
TP391
- DOI:
-
10.11992/tis.202203051
- Abstract:
-
Multi-view clustering is a research hotspot in the field of unsupervised learning. Of the many excellent multi-view clustering studies that have recently arisen, most assume that each view is complete. However, in a real scene, the data are extremely easily missed in the collection process, resulting in partially incomplete views. Simultaneously, many methods use traditional machine learning, i.e., the shallow-layer model, to learn data features, which makes it difficult for the model to mine the complex information of high-dimensional data. To solve these problems, in this paper, a novel deep mutual information maximization method is proposed for incomplete multi-view clustering. First, a deep autoencoder is used to learn the rich complex information of each view, and the knowledge of consistency among views is learned by the mutual information between potential representations. Then, the missing data are fixed up by the common latent representation of multi-view data. Additionally, this paper uses a self-paced strategy to fine-tune the model as it learns the samples from easy to difficult, obtaining a more clustering-friendly representation. Experiments performed on several real datasets show the effectiveness of our proposed method.