This repository aims at testing and comparing several imputation methods before performing LDA classification.

We run our experiments on 2 datasets. The first one is a data banknote authentification dataset containing four numerical features and a binary class. The second is simply generated by sampling from 2 multivariate gaussian distributions. The nature of the covariance matrices as well as the dimension (number of features) can be defined via command line arguments, as well as the missingness probability and the type of missingness. The imputation methods that get tested are "Grand Mean", "Conditional Mean", "Nearest Neighbour", "Regression". The concept of conditional mean is described in the pdf file at the root of the repository, along with the description of my experiments and conclusion.

matplotlib=3.8
numpy=1.26
pillow=10.0
python=3.12
scikit-learn=1.3.0

cd ImputationForLDA
conda create --name {env} --file requirements.txt

python main.py --dimensions 5 --cov_matrice normal --probs_missingness 0.1 --type_missingness MCAR

Please reach out.

Mathieu Charbonnel

• December 2023
  • Initial Release

This project is not licensed.

To Robert J. Durrant who supervised this scientific work.
Lohweg,Volker. (2013). banknote authentication. UCI Machine Learning Repository.