VIGAN: Missing View Imputation with Generative Adversarial Networks.

Area of Science:

• Data Science
• Bioinformatics
• Machine Learning

Background:

• Big data presents challenges in data quality and completeness, especially with multi-view or multi-modal datasets.
• The missing view problem, where entire data views are absent for samples, is a significant hurdle in multi-view analysis.
• Traditional methods like imputation or sample removal are often ineffective or reduce statistical power.

Purpose of the Study:

• To propose a novel approach for view imputation in multi-view datasets.
• To address the limitations of existing methods in handling missing data views.
• To improve the statistical power and analytical capabilities of multi-view data analysis.

Main Methods:

• Developed VIGAN, a novel approach utilizing generative adversarial networks (GANs) for view imputation.
• Treated each data view as a separate domain, identifying domain-to-domain mappings via GANs.
• Employed a multi-modal denoising autoencoder (DAE) to reconstruct missing views using GAN outputs and paired data.

Main Results:

• VIGAN demonstrated effective recovery of missing views through joint optimization of GAN and DAE.
• Empirical results on benchmark datasets validated VIGAN's superior performance against state-of-the-art methods.
• The approach proved effective and usable in a genetic study of substance use disorders.

Conclusions:

• VIGAN offers a powerful solution for the missing view problem in multi-view data analysis.
• The method enhances data completeness and statistical power, enabling more robust downstream analyses.
• VIGAN shows significant promise for applications in life sciences and other fields dealing with complex multi-modal data.