This repository compares adapted linear classifiers for handling missing values, specifically Perceptron, Logistic Regression, and Linear Discriminant Analysis (LDA). Two main approaches are considered for adapting to missing data: imputation methods and pattern-by-pattern methods.

• Constant imputation: Replacing missing values with a fixed value (e.g., 0).
• Iterative imputation: Using IterativeImputer from Scikit-learn to estimate missing values.

• These methods decompose the Bayes classifier for each missingness pattern, training a separate classifier for each observed pattern.
• In the LDA setting, this decomposition allows leveraging distributional assumptions:
  • When missingness is not informative, it is possible to average over all observed data.
  • When missingness is informative, it is possible to only estimate the most frequently observed patterns.

Multiple experiments are conducted to analyze:

• Convergence rates and robustness to misspecifications in the joint distribution and missingness model.
• The impact of correlation between input covariates.
• The effects of signal-to-noise ratio, missingness probability, and dimensionality.

For more details, refer to the paper:
https://arxiv.org/abs/2405.09196