Exploring the Effect of Model Assumptions on Prediction Performance of Bayesian Networks

Exploring the Effect of Model Assumptions on Prediction Performance of Bayesian Networks

Goslings, Wessel (TU Delft Electrical Engineering, Mathematics and Computer Science)

Nane, G.F. (mentor)

Delft University of Technology

Applied Mathematics

2021-02-06

This thesis concerns itself with the effect of the normality assumption, the effects of discretisation choices and other assumptions made by software on prediction performance, when using Gaussian Bayesian Networks. To test these effects, different types of Bayesian Networks are constructed and made to perform predictions, using the same dataset. The dataset used contains records regarding the citations and other bibliometric statistics of articles published by authors aliated with the Delft University of Technology between 2010 and 2014. The first model is a Gaussian Bayesian Network (GBN), which assumes that the conditional probability distributions (CPD) of all variables concerned are Gaussian. The second model is a Multinomial Bayesian Network (MBN), which uses discrete variables. To accommodate to this model, the data is discretized. The third model is the Hybrid Bayesian Network (HBN, which can handle both discrete and continuous data and has no normality assumption on the distribution of the variables. The last model is a non-parametric Bayesian Network (NPBN). To compare the different models, they are used to perform a set of predictions in the form of quantile estimation. The results show that the GBN performs as well as the NPBN, when looking at the bulk of the data. When looking at data, where the mean citation score (mcs), the predicted variable, exceeds the 75-% quantile, the performance of the GBN becomes much worse than that of the NPBN.

Bayesian Network
Structure learning
Citation
Normality Assumption

http://resolver.tudelft.nl/uuid:578a26af-1b80-44a2-b7f2-543f57521280

Student theses

bachelor thesis

© 2021 Wessel Goslings