Generative Imputation and Stochastic Prediction.

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Incomplete datasets are prevalent in machine learning applications.
• Existing missing data imputation techniques primarily focus on filling values, often neglecting associated uncertainties.
• Missing values introduce uncertainty in both data distribution and target class assignments.

Purpose of the Study:

• To propose a simple and effective method for imputing missing features in incomplete datasets.
• To develop a technique for estimating the distribution of target assignments given incomplete data.
• To address classification uncertainties arising from missing data.

Main Methods:

• A generator network is trained to create imputations for missing features.
• A discriminator network is employed to distinguish generated imputations from real data.
• A predictor network is trained on imputed samples to capture and utilize classification uncertainties for predictions.

Main Results:

• The proposed method demonstrates effectiveness in generating accurate imputations for missing data.
• The approach successfully estimates class uncertainties in classification tasks with missing values.
• Evaluations on image (CIFAR-10, MNIST) and tabular datasets show robust performance across various missingness rates.

Conclusions:

• The developed method provides a robust solution for handling incomplete datasets in machine learning.
• The technique effectively addresses both feature imputation and classification uncertainty estimation.
• This approach enhances predictive accuracy and reliability when dealing with missing data.