Learning a Bayesian Network from Forest Fire Data

Abstract

Bayesian networks are powerful tools for performing inference on data. Because of the growing complexity of the data being used, creating accurate Bayesian network structures is becoming more time-consuming and more dependant on expert knowledge. Investigations have been conducted into "structure learning": learning Bayesian network structures from data. Several different structure learning algorithms have been developed, each using different methods. In our research, we used two such algorithms -- SI HITON-PC and tabu search -- to learn Bayesian network structures from data consisting of measurements of forest fires. We investigated the impact of some of the algorithm parameters (the $\alpha$ parameter of SI HITON-PC and the scoring function of tabu search) on the produced results, and compared the best networks produced by both algorithms to one another and to a manually constructed network. These comparisons were made using measures of betweenness centrality and degree centrality, as well as Structural Hamming Distance (SHD). We found that even though the two algorithms we tested use different approaches, the networks they produce are quite similar, while both outputs differ noticeably from the manually constructed network.

Description

This repository contains the code of the second assignment for the NWI-IMC012 Bayesian Networks course at Radboud University (2019-2020). This assignment builds upon the previous assignment.

Structure

This project has the following file structure:

• preprocessing.R: contains the code that preprocesses forestfires.csv to obtain ff_preprocessed.csv (see data/).
• tools.R: contains helper functions for the analysis in analysis.R.
• analysis.R: contains the code for performing the analysis conducted in this assignment.
• BN_Assignment2.Rproj: the project settings as used by RStudio.
• data/: directory containing the original dataset forestfires.csv, obtained from the UCI Machine Learning Repository, and ff_preprocessed.csv, obtained by running preprocessing.R. The latter is the same data which was used for the previous assignment of this course
• img/: directory that contains all images generated by preprocessing.R.

Contributors

Freek van den Bergh
Max Driessen
Xiaoxuan Lei