Learning Bayesian Networks: The Combination of Scoring Function and Dataset
Saratha Sathasivam1, Poh Choo Song2, Jing Jie Yeap3

1Saratha Sathasivam, School of Mathematical Sciences, Universiti Sains Malaysia, 11800 USM, Penang, Malaysia. 
2Poh Choo Song*, Faculty of Science, Universiti Tunku Abdul Rahman, Perak, Malaysia.
3Jing Jie Yeap, Faculty of Science, Universiti Tunku Abdul Rahman, Perak, Malaysia. 
Manuscript received on May 05, 2020. | Revised Manuscript received on May 15, 2020. | Manuscript published on June 30, 2020. | PP: 149-154 | Volume-9 Issue-5, June 2020. | Retrieval Number: D7568049420/2020©BEIESP | DOI: 10.35940/ijeat.D7568.069520
Open Access | Ethics and Policies | Cite | Mendeley
© The Authors. Blue Eyes Intelligence Engineering and Sciences Publication (BEIESP). This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)

Abstract: Bayesian network (BN), a graphical model consists nodes and directed edges, which representing random variables and relationship of the corresponding random variables, respectively. The main study of Bayesian network is structural learning and parameter learning. There are score-and-search based, constraint based and hybrid based in forming the network structure. However, there are many types of scores and algorithms available in the structural learning of Bayesian network. Hence, the objective of this study is to determine the best combination of scores and algorithms for various types of datasets. Besides, the convergence of time in forming the BN structure with datasets of different sizes has been examined. Lastly, a comparison between score-and-search based and constraint based methods is made in this study. At the end of this study, it has been observed that Tabu search has the best combination with the scoring function regardless of the size of dataset. Furthermore, it has been found that when the dataset is large, the time it takes for a BN structure to converge is shorter. Last but not least, results showed that the score-and-search based algorithm performs better as compared to constraint based algorithm.
Keywords: Bayesian network, belief network, convergence, score, structural learning.