Webinar in Bayesian Networks

[michael tsagris]

The Department of Economics at the University of Crete welcomes you to attend a webinar on Wednesday 12 November at 15.00 GREEK time.

Title: Investigation of the Hyper-parameters and the Scoring Functions of PCTABU and PCHC Bayesian Network Algorithms

Presenter: Sevinç Volkan, Muğla Sıtkı Koçman University (Turkey)

Abstract: This seminar has three main goals. The first one is to introduce a novel Bayesian network learning algorithm called PCTABU. The second one is to examine the effects of hyper-parameter choices on Bayesian networks by comparing the estimation performances of the newly introduced algorithm and a previously suggested algorithm PCHC, which use the hyper-parameters Tabu-Search and Hill-Climbing, respectively. The third goal, on the other hand, is to investigate the effects of scoring function choice on model performances by making the aforementioned comparison using three different scoring functions Bayesian Dirichlet equivalence, log-likelihood, and Bayesian information criterion. The comparison was performed by using both simulated and real-data-based Bayesian networks. The results show that both hyper-parameter and scoring function choices have significant effects on the performance of Bayesian networks, such that the newly introduced PCTABU algorithm appears to perform better for the networks with lower Markov blanket scores, average degrees, and densities. Another result is that the choice between Hill-Climbing and Tabu-Search hyper-parameters may affect the performances of Bayesian network algorithms. In addition, the log-likelihood scoring function should be avoided, as it produced the worst results most of the time. Bayesian Dirichlet equivalence and Bayesian information criterion, on the other hand, seem to produce similar results.

Zoom link: https://uoc-gr.zoom.us/j/88659969718?pwd=g6bjYPDCuUQo1bzVxjjbgQL4xFN1f3.1