Probabilistic graphical models : bayesian networks

• Main Author: Chan, Xiang Yun
• Other Authors: Frederique Elise Oggier
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2021
• Subjects: Science::Mathematics::Statistics; Science::Mathematics::Analysis
• Online Access: https://hdl.handle.net/10356/148509
• Institution: Nanyang Technological University

This report conducts a review on Bayesian networks and provides a Bayesian network analysis of a dataset provided by Hewlett-Packard. The main focus of the review is on discrete Bayesian networks while there is a brief mention of Gaussian and hybrid Bayesian networks. Methodologies on constructing Bayesian networks are included. Exploratory data analysis was conducted on the dataset before experimenting with the types of Bayesian networks that can be constructed. As the dataset contains a mix of discrete and continuous variables, few discretization methods were used on the continuous variables to produce a discrete Bayesian network. Gaussian and hybrid networks were generated as well. As a result, there were many preliminary structures to choose from. As the objective of building a network is to model the relationships among variables, structures with disjoint variables were eliminated. Six network structures were narrowed down and evaluated using network scores, expected log-loss and mean-squared errors in predictions. Inference was performed for the variables of interest.