This repository contains the experiments for "Semiparametric Bayesian Networks."

There is a folder for each experiment type. synthetic for synthetic data experiments, GaussianNetworks for data sampled from Gaussian networks and UCI data for experiments from the UCI repository.

To run these experiments PyBNesian is needed.

To ensure the reproducibility of the experiments, we used the following experimental setup:

• The experiments were run using this version of PyBNesian. We believe that the results should be reproducible with other versions.

• Ubuntu 16.04, Python 3.7.3, and compiled with gcc 9.3.0. We believe that the results should be reproducible for most configurations. Altough sometimes C++ compilers introduce optimizations and changes in the standard library implementation in new releases. These changes may subtly affect the results.

Organization

All experiments contain scripts with the following structure:

• experiments_helper.py
• train_[algorithm]_[model_type].py
• test_[algorithm]_[model_type].py
• other code such as plotting scripts.

experiments_helper.py defines the parameters of the experiment at the start of the file. Also, it contains some auxiliary code used by the experiments.

The experiments are executed in two steps so the experiments can be paused/resumed easily. First, the models are learned from the training data and saved. For this, use the scripts train_[algorithm]_[model_type].py. This step can take quite some time, so the execution can be stopped at any moment for all training scripts. If the script is executed again, it will automatically detect the already learned models.

Then, the test_[algorithm]_[model_type].py scripts load the learned models and test them on unseen data. The results of this testing is saved in csv files or printed in the screen depending on the type of experiment.

Train scripts

All the learned models are saved in a local folder called models/.

Important: create a models/ folder on each experiment folder.

All the learned models (including all the iterations of the greedy hill-climbing algorithm) are saved.

The [algorithm] part can take the following values:

• hc for greedy hill-climbing.
• pc for PC.

The [model_type] part can take the following values:

• graph: This option is only available for PC algorithm. It learns the graph of the Bayesian network. Important: when learning models with PC, this script should be run before learning the specific Bayesian networks described below.
• gbn to learn Gaussian Networks.
• kdebn to learn kernel density estimation Bayesian networks.
• spbn to learn semiparametric Bayesian networks with linear Gaussian CPDs as starting node types.
• spbn_ckde to learn semiparametric Bayesian networks with conditional kernel density estimation CPDs as starting node types.

Test scripts

The structure of the test scripts is the same as in the train scripts. This process usually takes much less time than the training. For the UCI data experiment, the scripts save the results in CSV files called results_[algorithm]_[model_type].csv. For the other experiment types, the experiments are printed on screen.

Other scripts

The remaining scripts contain useful utilities such as plotting, sampling of data or statistical tests.

Execution Times

For the execution times experiment, the data must be first generated calling generate_dataset_spbn.py. The figure in the article is generated by plot_times.py. The models must be trained first by calling all train_[algorithm]_[model_type].py scripts.

Bearings Degradation

The figures in the article are generated by plot_bearings.py.