Comprehensive Multiview Representation Learning via Deep Autoencoder-Like Nonnegative Matrix Factorization.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Multiview representation learning (MRL) is vital for analyzing complex data from multiple sources.
• Existing Nonnegative Matrix Factorization (NMF)-based MRL methods are often shallow and neglect hierarchical information.
• Deep Matrix Factorization (DMF) methods exist but focus on view consistency and involve complex clustering.

Purpose of the Study:

• To propose a novel Deep Autoencoder-like NMF for Multiview Representation Learning (DANMF-MRL) to address limitations of existing MRL techniques.
• To develop a unified one-step DANMF-MRL model for simultaneous latent representation learning and clustering.
• To enhance multiview representation by considering both consistency and complementarity.

Main Methods:

• Developed a Deep Autoencoder-like NMF for Multiview Representation Learning (DANMF-MRL) framework.
• Introduced a one-step DANMF-MRL model integrating representation learning and clustering.
• Designed efficient iterative optimization algorithms with theoretical convergence analysis.

Main Results:

• The DANMF-MRL framework effectively captures comprehensive multiview representations by considering consistency and complementarity.
• The one-step DANMF-MRL achieved optimal clustering performance by unifying representation and clustering.
• Experimental results on five benchmark datasets demonstrated the superiority of the proposed methods over state-of-the-art MRL techniques.

Conclusions:

• The proposed DANMF-MRL models offer a significant advancement in multiview representation learning.
• The one-step approach simplifies the process and improves clustering accuracy.
• These methods provide a powerful tool for analyzing complex, multi-source data.