Adaptive Thresholding in Structure Learning of a Bayesian Network英文资料.pdf

Proceedings of the Twenty-Third International Joint Conference on Artificial Intelligence Adaptive Thresholding in Structure Learning of a Bayesian Network Boaz Lerner, Michal Afek, Rafi Bojmel Ben-Gurion University of the Negev, Israel boaz@bgu.ac.il; {caspimic,rafibojmel}@ Abstract threshold selection – a wrong threshold influences structure Thresholding a measure in conditional learning [2, 12, 14, 20]. Commonly, a threshold is selected independence (CI) tests using a fixed value enables according to three main methods: as a user default [10, 29, learning and removing edges as part of learning a 31, 33], based on limited trial and error experimentation [5], Bayesian network structure. However, the learned or automatically as the threshold, from a range of structure is sensitive to the threshold that is candidates, that max(in)imizes a performance measure [9, commonly selected: 1) arbitrarily; 2) irrespective 14, 20, 23, 35]. The first method provides simplicity and of characteristics of the domain; and 3) fixed for all low runtime, but cannot guarantee satisfactory performance CI tests. We analyze the impact on mutual for most problems. The second method overcomes information – a CI measure – of factors, such as exhaustive experimentation using a limited number of sample size, degree of vari