作者: David D. Jensen , Andrew S. Fast
DOI:
关键词:
摘要: A Bayesian network is graphical representation of the probabilistic relationships among set variables and can be used to encode expert knowledge about uncertain domains. The structure this model represents conditional independencies in data. networks are widely applicable, having been domains ranging from monitoring patients an emergency room predicting severity hailstorms. In thesis, I focus on problem learning Under certain assumptions, learned a represent causal data. Constraint-based algorithms for designed accurately identify distribution underlying data and, therefore, relationships. These use series hypothesis tests learn independence constraints model. When sample size limited, these prone errors. present comprehensive empirical evaluation constraint-based show that existing many false negative errors due running with low statistical power. Furthermore, analysis shows solutions fail reduce overall algorithms. I new inspired by constraint satisfaction able produce significant improvements structural accuracy. exploit interaction error. First, introduce algorithm based optimization sound limit, like algorithms, but guaranteed DAG. This learns models accuracy equivalent or better algorithms. Second, relaxation. Constraint relaxation combines different techniques likely incorrect, remove those consideration. combining produces significantly when compared demonstrating effectiveness approaches accurate networks.