Deeply-Learned Generalized Linear Models with Missing Data.

Area of Science:

• Machine Learning
• Statistical Modeling
• Data Science

Background:

• Deep Learning (DL) methods are increasingly popular for supervised learning but struggle with complex missing data.
• Existing DL models face challenges with ignorable and non-ignorable missingness patterns in datasets.
• Handling missing data is crucial for the reliable application of DL in real-world scenarios.

Purpose of the Study:

• To formally treat missing data within deeply learned generalized linear models (DLGLMs).
• To introduce a new DL architecture, dlglm, capable of handling both ignorable and non-ignorable missingness.
• To evaluate the performance of dlglm against existing methods, particularly with missing not at random (MNAR) data.

Main Methods:

• Developed a novel deep learning architecture, dlglm, for regression and classification.
• Implemented flexible accounting for ignorable and non-ignorable missing data patterns during training.
• Utilized statistical simulations to compare dlglm performance with existing supervised learning approaches.

Main Results:

• The proposed dlglm architecture demonstrates superior performance in supervised learning tasks with MNAR missingness.
• Statistical simulations confirm the effectiveness of dlglm in handling complex missing data scenarios.
• The method shows promise for real-world applications, as evidenced by a case study.

Conclusions:

• dlglm offers a robust solution for supervised learning problems with complex missing data.
• The architecture effectively addresses limitations of existing DL methods when dealing with MNAR data.
• This work advances the application of DL in scenarios with incomplete datasets.