This repository contains the code to reproduce the experiments of the paper:
NeuMiss networks: differential programming for supervised learning with missing values.
https://arxiv.org/abs/2007.01627
The file NeuMiss.yml indicates the packages required as well as the versions used in our experiments.
The methods used are implemented in the following files:
- neumannS0_mlp: the NeuMiss network.
- mlp: the feedforward neural network.
- estimators: the other methods used.
The files ground_truth and amputation contain the code for data simulation and the code for the Bayes predictors.
To reproduce the experiments, use:
python launch_simu_perf.py MCARpython launch_simu_perf.py MAR_logisticpython launch_simu_perf.py gaussian_smpython launch_simu_perf.py probit_smpython launch_simu_depth_effect.pypython launch_simu_archi.py
These scripts save their results as csv files in the results foder. The plots can be obtained from these csv files by running the plots_xxx files.
All datasets are found in the datasets folder.
diabetes.csv: https://www.kaggle.com/mathchi/diabetes-data-set/version/1- Contains missing values
diabetes_complete.csv: https://www.kaggle.com/homayoonkhadivi/ai-for-medical-prognosis-diabetes-datasets?select=y_data.csv- Complete dataset (no missing values)