Incomplete multi-view clustering network via nonlinear manifold embedding and probability-induced loss

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Incomplete multi-view clustering presents significant challenges due to missing data across different views.
• Existing methods struggle with redundant features, lack of local structure consideration, and inadequate handling of outliers and noise.
• Addressing these limitations is crucial for advancing clustering techniques in complex datasets.

Purpose of the Study:

• To propose a novel deep clustering network for incomplete multi-view data.
• To overcome the limitations of existing methods by learning discriminative features, incorporating local structure, and robustly handling data distribution.
• To enhance the accuracy and effectiveness of clustering algorithms on datasets with missing information.

Main Methods:

• A deep clustering network combining multi-view autoencoders with Uniform Manifold Approximation and Projection (UMAP) for latent feature extraction.
• Integration of Gaussian Mixture Model (GMM) to model complex data distributions and mitigate outlier effects.
• Development of a probability-induced loss function to jointly optimize feature learning and clustering within a unified framework.

Main Results:

• The proposed method effectively extracts latent consistent features from incomplete multi-view data.
• Experimental results on benchmark datasets demonstrate superior performance compared to existing approaches.
• The network successfully addresses issues of redundant features, local structure, and data distribution for robust clustering.

Conclusions:

• The novel deep clustering network provides an effective solution for incomplete multi-view clustering.
• The integration of UMAP, GMM, and a probability-induced loss function significantly improves clustering performance.
• This approach offers a robust and efficient method for analyzing complex datasets with missing data across multiple views.