What to do when you don't have much data: Issues in small-sample parameter learning in Bayesian networks.

摘要

A discrete Bayesian network is a factorization of a joint distribution over random variables. The most common use of these networks is for the computation of conditional probabilities (query responses). The parameters of these networks are often learned from data. Thus network parameters are themselves uncertain, which induces a distribution for any query response. When data sets are small, the effects of parameter uncertainty can be severe. In this thesis we argue that when data sets are small, the distribution of a query response is accurately modelled by a Beta distribution. Procedures for the modeling of the query response are also reviewed. Furthermore, we examine proposed techniques for parameter learning when data sets are small and only one query is of interest.