Grouped Feature Importance and Combined Features Dependence Plot

This repository gives access to an implementation of the methods presented in the paper submission “Grouped Feature Importance and Combined Features Dependence Plot”, as well as all code that was used for the simulations and the usecase.

This repository is structured as follows:

    ├── code/                     # All code               
    |   ├── functions/            # Implemented functions
    |   ├── simulations/          # Scripts for simulation examples in paper
    |   |   ├── analysis/         # Scripts used to create figures and tables in the paper for simulation examples
    |   |   ├── experiments/      # Scripts used to create data for simulation examples
    |   ├── usecase/              # Scripts used to create data, figures and tables in usecase section
    ├── results/                  # Location where all data and figures are stored
    │   ├── figures/              # Location where final figures for paper are stored
    │   ├── simulation_results/   # Location where all data of simulation examples is stored
    │   ├── usecase_results/      # Location where all usecase data is stored
    ├── LICENSE
    └── README.md               

Reproduce Experiments

Steps to reproduce the experiments of the simulation experiments.

1. Install all required packages.

# from CRAN
install.packages(c("ranger", "dplyr", "batchtools", "mlr", "ggplot2", "gridExtra", "tidyr", "reshape2",
"ggExtra", "future.apply", "BBmisc", "data.table", "stringi", "checkmate", "kernlab", "xtable", "mlrCPO", "devtools", "PMA"))

# from github
devtools::install_github("giuseppec/featureImportance")
install_github("nickseedorff/totalvis")

Alternatively, we also provide a custom docker image for this paper. First, install docker and then build the image in your terminal (from the projects directory). The Dockerfile is provided.

docker build -t gimp .

Run the image with:

docker run --rm -p 8787:8787 -e PASSWORD=11111 gimp

Open localhost:8787 in your browser. Login with the username rstudio and the password 11111. You can then run all code in this RStudio server instance.

2. Create an experimental registry, add experiments and problem and run simulations via scripts in subfolder code/simulations/experiments. Data produced by the scripts is stored in the subfolder results/simulation_results as a separate registry.

3. To reproduce figures and tables of Section 3 and 4, run the scripts in subfolder code/simulations/analysis. The scripts are names with respect to the references subsections in the paper. Figures produced within the script are stored in results/figures the subfolder results/simulation_results as a separate registry.

4. To reproduce figures and tables of Section 5, run the scripts in subfolder code/usecase:

    code/usecase/psychology_feat_imp.R                     # Code for feature importance measures               
    code/usecase/psychology_gfs.R                          # Code for sequential grouped feature importance
    code/usecase/psychology_sspca.R                        # Code for CFEP
    code/usecase/psychology_collect_results.R              # Code for tables and figures

Data for the usecase is from https://osf.io/kqjhr/ and is available as RData file in this repository.