Empirical Behavior of Bayesian Network Structure Learning Algorithms

Abstract

Bayesian network structure learning BNSL is the problem of finding a BN structure which best explains a dataset. Score-based learning assigns a score to each network structure. The goal is to find the structure which optimizes the score. We review two recent studies of empirical behavior of BNSL algorithms. The score typically reflects fit to a training dataset; however, models which fit training data well may generalize poorly. Thus, it is not clear that finding an optimal network is worthwhile. We review a comparison of exact and approximate search techniques. Sometimes, approximate algorithms suffice; for complex datasets, the optimal algorithms produce better networks. BNSL is known to be NP-hard, so exact solvers prune the search space using heuristics. We next review problem-dependent characteristics which affect their efficacy. Empirical results show that machine learning techniques based on these characteristics can often be used to accurately predict the algorithms' running times.