Ioannis Tsamardinos, Laura E. Brown
Modern Bayesian Network learning algorithms are time-efficient, scalable and produce high-quality models; these algorithms feature prominently in decision support model development, variable selection, and causal discovery. The quality of the models, however, has often only been empirically evaluated; the available theoretical results typically guarantee asymptotic correctness (consistency) of the algorithms. This paper describes theoretical bounds on the quality of a fundamental Bayesian Network local-learning task in the finite sample using theories for controlling the False Discovery Rate. The behavior of the derived bounds is investigated across various problem and algorithm parameters. Empirical results support the theory which has immediate ramifications in the design of new algorithms for Bayesian Network learning, variable selection and causal discovery.
Subjects: 9.1 Causality; 12. Machine Learning and Discovery
Submitted: Apr 15, 2008