Modelling citation performance of Canadian researchers using Bayesian Networks

Modelling citation performance of Canadian researchers using Bayesian Networks

Stuurman, Pim (TU Delft Electrical Engineering, Mathematics and Computer Science)

Nane, Tina (mentor)

Delft University of Technology

2017-08-22

This thesis is about modelling citation performance using Bayesian networks (BN). As an approximation for citation performance, we will use the proportion of papers by an author that resulted in being one of the top 10% most cited papers. Using 15 predicting variables such as year of first publication and average number of authors per paper, we analyze dependence structure and predictive power with BNs.

We compare different continuous (CBN) and non-parametric (NPBN) Bayesian network models on predictive performance and consistency. Assumptions, structure learning algorithms and prediction methods for the different models are contrasted with those for linear regression and k-nearest neighbour models. A simulated dataset from a Gaussian copula is used to show the predictive power of NPBNs, and some of the challenges in consistency for CBNs.

Furthermore, we introduce a new prediction method for NPBNs, applying a k-nearest neighbour model to a dataset generated from the network. While overall predictive performance of this method is worse than conditionalization, it has some computational benets and outperforms conditionalization in extreme values (95th percentile) on a simulated dataset.

The main conclusion of this thesis is that CBN and NPBN models perform similar to linear regression models in predicting citation performance of Canadian researchers. However, none of the Bayesian networks and linear regression models fit perfectly, as they all have violations of their model assumptions.

Continuous Bayesian Networks
non-parametric Bayesian Networks
Bayesian networks
Bayesian network structure learning
citation analysis
bibliometric studies

http://resolver.tudelft.nl/uuid:d702ba0b-e568-485c-b2e2-0ac483196a53

Student theses

master thesis

© 2017 Pim Stuurman