Multiple Imputation via Generative Adversarial Network for High-dimensional Blockwise Missing Value Problems.

Area of Science:

• Computer Science
• Statistics
• Machine Learning

Background:

• Missing data is a common issue in real-world datasets, impacting prediction accuracy and statistical consistency.
• Multiple Imputation (MI) is the standard for handling missing data, accounting for uncertainty and enabling valid statistical inference.

Purpose of the Study:

• To introduce Multiple Imputation via Generative Adversarial Network (MI-GAN), a novel deep learning approach for handling missing data.
• To demonstrate MI-GAN's effectiveness under the Missing At Random (MAR) mechanism with theoretical validation.

Main Methods:

• Developed MI-GAN, a Generative Adversarial Network (GAN)-based multiple imputation technique.
• Utilized conditional generative adversarial neural networks for imputation.
• Validated the method on high-dimensional datasets.

Main Results:

• MI-GAN achieves imputation error comparable to state-of-the-art methods on high-dimensional data.
• Demonstrated superior performance in statistical inference compared to existing imputation techniques.
• Showcased significant improvements in computational speed.

Conclusions:

• MI-GAN offers a powerful, deep learning-based solution for multiple imputation.
• The method provides accurate imputation and robust statistical inference, with enhanced computational efficiency.