Evaluating the impact of multivariate imputation by MICE in feature selection

Area of Science:

• Machine Learning
• Data Preprocessing
• Statistical Analysis

Background:

• Handling missing values is critical in machine learning.
• Traditional methods like discarding data or using mean/median imputation can introduce bias.
• Most machine learning algorithms require complete datasets for analysis.

Purpose of the Study:

• To demonstrate the positive impact of multivariate imputation on feature selection for datasets with missing values.
• To compare multivariate imputation against basic imputation techniques and complete case analysis.

Main Methods:

• Feature selection was performed on complete datasets, incomplete datasets (5-50% missingness), and datasets imputed using basic techniques and multivariate imputation (MICE).
• Standard feature selection algorithms were employed for comparison.

Main Results:

• Datasets imputed using multivariate imputation (MICE) yielded superior results in the feature selection process.
• Multivariate imputation outperformed basic imputation methods and non-imputed incomplete datasets.

Conclusions:

• Applying multivariate imputation, specifically MICE, effectively reduces bias in the feature selection process.
• Multivariate imputation is a recommended strategy for handling missing data in machine learning workflows.