Realize Generative Yet Complete Latent Representation for Incomplete Multi-View Learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Multi-view learning environments often suffer from missing observations due to inherent limitations.
• Existing representation learning methods struggle to fully utilize information, lacking either cross-generative (filling missing data) or solidative (consistent representation inference) capabilities.

Purpose of the Study:

• To propose a deep generative model, Complete Multi-view Variational Auto-Encoders (CMVAE), for learning complete generative latent representations from incomplete multi-view data.
• To address the challenge of missing observations by modeling view generation from a complete latent variable and resolving missing views via posterior distribution estimation.

Main Methods:

• Developed CMVAE, a deep generative model utilizing a mixture of Gaussian distributions for complete latent variables.
• Introduced a novel variational lower bound to integrate view-invariant information, enhancing the solidity of learned representations.
• Employed techniques to mine intrinsic correlations between views for cross-view generality and fused information using view weights for solidity.

Main Results:

• CMVAE demonstrated superior performance on benchmark tasks including clustering, classification, and cross-view image generation.
• Analyses confirmed the efficiency and robustness of CMVAE regarding time complexity and parameter sensitivity.
• The model's practical significance was exemplified through its application to bioinformatics data.

Conclusions:

• CMVAE effectively learns complete generative latent representations from incomplete multi-view data.
• The proposed method offers a robust and efficient solution for multi-view learning challenges, outperforming existing approaches.
• CMVAE shows significant promise for applications in diverse fields, including bioinformatics.