Towards a unified framework for graph-based multi-view clustering | JoVE Visualize

Area of Science:

Data Science
Machine Learning
Computer Vision

Background:

Multi-view clustering is crucial for real-world applications, with common methods including spectral clustering, subspace methods, matrix factorization, and kernel methods.
Existing approaches often fuse similarity matrices directly, making them susceptible to noise and separating affinity learning from clustering.
This limitation can degrade performance when dealing with noisy data across multiple views.

Purpose of the Study:

To propose a novel method, One Step Multi-view Clustering via Consensus Graph Learning and Nonnegative Embedding (OSMGNE), to address the limitations of existing multi-view clustering techniques.
To develop a method that robustly handles noisy similarity matrices by learning a consensus similarity matrix.
To enable simultaneous estimation of multiple matrices and automatic view weighting without hyperparameters.

Main Methods:

Introduced the One Step Multi-view Clustering via Consensus Graph Learning and Nonnegative Embedding (OSMGNE) method.
Developed a consensus graph learning approach to mitigate noise in individual view similarity matrices.
Incorporated nonnegative embedding for direct soft cluster assignment, eliminating post-processing steps.
Proposed an iterative algorithm to solve the optimization problems for two variations of the method.

Main Results:

The proposed OSMGNE method effectively learns a consensus similarity matrix, reducing the impact of noisy data.
The nonnegative embedding allows for direct generation of cluster assignments, simplifying the process.
The method jointly estimates multiple matrices (similarity, spectral projection, indicator) and automatically determines view weights.
Experimental results on real datasets demonstrate the superiority of the proposed methods over existing approaches.

Conclusions:

The novel OSMGNE method offers a robust and efficient solution for multi-view clustering, particularly in the presence of noisy data.
Jointly learning consensus representations and cluster assignments improves overall clustering accuracy and stability.
The method's ability to handle multiple subtasks simultaneously and avoid hyperparameters makes it a practical advancement.