EDNEL: Estimation of Dependency Networks Algorithm for Ensemble Learning

Installation

On Windows (tested on Windows 10):

1. Download Java JDK SE version 8u261 from Oracle site: https://www.oracle.com/br/java/technologies/javase/javase8u211-later-archive-downloads.html (note: you'll need to have an Oracle account to do so, to accept their software use agreement)
2. Install using the Graphical Installer Note: installing JDK on windows should have already installed the Java Runtime Environment (JRE). If not, download version 8u241 here: https://www.java.com/en/download/win10.jsp
3. Test whether these two programs are in PATH variable:

   java -version
   javac -version

If any of the programs was not found, add their installation path to the PATH system variable (the installation path is the one selected in steps 1 and 2)

Note: the code uses JDK version 8, independent of the downloaded version.

On Ubuntu (tested version 18.04):

You still have to download JDK from Oracle. On Ubuntu this is a little more complicated because you have to use a third party tool to download it. Follow these steps:

1. Create an Oracle account
2. Download version 14.0.2 from this page: https://www.oracle.com/java/technologies/javase-jdk14-downloads.html Choose Linux Compressed Archive
3. Unpack it wherever it seems fit in your machine, e.g. in your home directory: tar -xf jdk-14.0.2_linux-x64_bin.tar.gz
4. A new folder of name jdk-14.0.2 was created in this folder. Add this path to your PATH variable: a. Open the .bashrc file in your HOME directory, e.g. nano .bashrc b. Pre-pend the path to jdk-14.0.2/bin in your PATH variable: export PATH=:/home/<your_user>/jdk-14.0.2/bin:$PATH c. You will have to either close and open the terminal or give a source .bashrc for modifications to have immediate effect c. Check whether everything is working:

   which java
   which javac

   It should point to your new installation path, as opposed to the system-wide java/javac.

Install Weka version 3.9.3 Download it from Source Forge: https://sourceforge.net/projects/weka/files/weka-3-9/3.9.3/

TODO add instructions from building code from command line

Using

After you have generated a .jar for the project, run it by invoking the following command:

java -jar ednel.jar

This will bring a list of parameters for the program. An successful call to the code is

java -jar ednel.jar --datasets_path keel_datasets_10fcv --datasets_names vehicle,pima,wisconsin,flare,australian,german,bupa,contraceptive --metadata_path metadata --n_generations 200 --n_individuals 200 --n_samples 1 --learning_rate 0.7 --selection_share 0.5 --burn_in 100 --thinning_factor 0 --early_stop_generations 200 --early_stop_tolerance 0.001 --log --max_parents 1 --delay_structure_learning 1 --n_jobs 5 

Interpreting data

Setup

Create a new Anaconda environment:

conda create --name ednel python=3.7 --yes
conda activate ednel
conda install --file requirements.txt
conda install -c alubbock pygraphviz=1.5 graphviz=2.41
conda install -c rmg pydot=1.4.1

Add the path to graphviz <folder_to_anaconda_installation>Anaconda3/pkgs/graphviz-2.41-0/Scripts to your system path

Generating predictive performance metrics metrics

---

If during the training process, flag --log was set, it will be possible to generate a report on the various predictive performance metrics available in Weka:

python scripts/postprocess.py --csv-path <path_to_run>

Example:

python scripts/postprocess.py --csv-path metadata/02-06-2020-09-00-18

Generating dashboard of training metadata

The same can be done for generating a visualization of the structure of evolutionary algorithm during training, as well as the explored solution space, probability tables, and population metrics:

python scripts/plot_network.py --experiment-path <path_to_run>

Example:

python scripts/plot_network.py --experiment-path "C:\Users\henry\Projects\trash\28-06-2020-14-14-49\dummygerman\sample_01_fold_01"

Testing compatibility of datasets

Use class TestDatasets to test whether EDNEL will successfuly run on a dataset:

java -classpath ednel.jar ednel.utils.analysis.tests.TestDatasets --datasets_path <datasets_path> --datasets_names null --metadata_path null --n_generations 2 --n_individuals 25 --n_samples 1 --learning_rate 0.7 --selection_share 0.5 --burn_in 2 --thinning_factor 0 --early_stop_generations 200 --early_stop_tolerance 0.001 --max_parents 0 --delay_structure_learning 1 --n_jobs 1 --timeout 300