[Paper Review] A Survey on Multi-View Clustering

Guoqing Chao, Shiliang Sun+1 more|arXiv (Cornell University)|Dec 18, 2017

Face and Expression Recognition134 references119 citations

TL;DR

This survey reviews and taxonomy-designs multi-view clustering (MVC) methods, separating generative and discriminative approaches and detailing five discriminative classes, relations to related paradigms, applications, and open problems.

ABSTRACT

With advances in information acquisition technologies, multi-view data become ubiquitous. Multi-view learning has thus become more and more popular in machine learning and data mining fields. Multi-view unsupervised or semi-supervised learning, such as co-training, co-regularization has gained considerable attention. Although recently, multi-view clustering (MVC) methods have been developed rapidly, there has not been a survey to summarize and analyze the current progress. Therefore, this paper reviews the common strategies for combining multiple views of data and based on this summary we propose a novel taxonomy of the MVC approaches. We further discuss the relationships between MVC and multi-view representation, ensemble clustering, multi-task clustering, multi-view supervised and semi-supervised learning. Several representative real-world applications are elaborated. To promote future development of MVC, we envision several open problems that may require further investigation and thorough examination.

Motivation & Objective

• Summarize the state of the art in multi-view clustering (MVC) and its motivations.
• Propose a taxonomy to organize MVC methods into meaningful categories.
• Discuss relationships between MVC and related learning paradigms (e.g., representation learning, ensemble clustering, multi-task learning).
• Highlight representative applications across domains and identify open problems for future work.

Proposed method

• Classify MVC methods into generative (model-based) and discriminative (similarity-based) approaches.
• Within discriminative MVC, further categorize into five classes based on how multiple views are combined: common eigenvector matrix, common coefficient matrix, common indicator matrix, direct view combination, and view combination after projection.
• Describe representative techniques such as multi-view spectral clustering, multi-view subspace clustering, multi-view NMF, multi-kernel clustering, and CCA-based methods.
• Explain how each class constructs and optimizes objective functions to achieve consistent clustering across views.
• Discuss the use of mixture models and EM in generative MVC, and explain CMMs and multi-view extensions using KL-divergence minimization across views.

Experimental results

Research questions

• RQ1 What are the main strategies used to combine multiple views in MVC?
• RQ2 How can MVC be categorized into coherent classes, and what are representative methods in each class?
• RQ3 What are the relationships between MVC and related areas such as multi-view representation, ensemble clustering, and multi-task learning?
• RQ4 What applications and open problems characterize the current MVC landscape?

Key findings

• MVC methods are largely discriminative, with a clear taxonomy that includes five classes of view combination.
• Generative MVC relies on mixture models and EM, including multi-view CMM extensions, to achieve clustering across views.
• There is a strong linkage between MVC and multi-view representation, ensemble clustering, multi-task clustering, and multi-view supervised/semi-supervised learning.
• MVC has been applied across domains such as computer vision, natural language processing, social multimedia, bioinformatics, and health informatics.
• The survey identifies open problems to guide future MVC development.