Benchmark of Multiple Imputation using Chained Equations (MICE) algorithm

This repo contains a benchmark of the MICE algorithm regarding performance and execution time for imputing missing values on data. For this purpose, several variations of the MICE algorithm have been implemented using LightGBM instead of linear models for value imputation for accuracy and speed improvements. Here, four method are evaluate against the mean / mode value imputation procedure for multi-dimensional data on 5 different datasets available on the Scikit-Learn package (namely, boston house prices, iris, diabetes, wine and breast cancer):

• Vanila MICE: Value by value imputation over a set number of iterations
• Fast MICE: Column by column imputation over a set number of iterations
• Slow-Fast MICE: Value by value imputation in the first iteration and column by column for the remaining iterations
• Fast-Slow MICE: Column by Column imputation in all iteration except the last one where value by value imputation is used for the remaining iterations

The procedure is available via a jupyter notebook in the notebook/ folder in this repo.

TL;DR

If you are just looking for the results of the benchmark, here they are:

• On average, Fast MICE is 12.0x faster than Fast-Slow / Slow-Fast MICE and 56.0x faster than Vanila MICE
• On average, Fast-Slow / Slow-Fast MICE are 5.0x faster than Vanila MICE

Boston house prices results

Iris results

Diabetes results

Wine results

Breast Cancer results

Requirements

• Python3 (3.6 recommended)
• jupyter
• scipy stack (pandas, scipy, scikit-learn, etc.)
• docker (optional, recommended)

Getting started

The code is available via jupyter notebooks for easier use.

To run these notebooks, you need to start a jupyter server. Here, you can do it in two ways:

• a) run a local jupyter server or
• b) run a self-contained docker image.

Run a local jupyter server

To start the jupyter server you must first have python + jupyter installed. The quickest way to accomplish this is by installing anaconda.

After installing anaconda, you should create an environment:

$ conda create -n py36_jupyter python=3.6 anaconda

This command will install the recommended version of CPython and the necessary packages to run the code.

Finally, to start a jupyter server you simply need to run the following command:

$ jupyter notebook

Run a self-contained docker image

To run the notebooks using docker, you first need to build the container's docker image. To do so, you just need to do the following:

• i) Build the container using a Makefile macro:

  $ make build

• ii) Run the container using a command:

  $ docker image build -t jupyter_scipy_custom .

Then, to start the container you can:

• i) Run the container using a Makefile macro:

  $ make run

• ii) Run the container using a command:

  $ docker run --rm -p 8888:8888 -v "$PWD"/notebook:/home/jovyan/work --name jupyter_benchmark_mice jupyter_scipy_custom

License

MIT