Bayesian Network Learning with Abstraction Hierarchies and Context-Specific Independence

Context-specific independence representations, such as tree-structured conditional probability tables (TCPTs), reduce the number of parameters in Bayesian networks by capturing local independence relationships and improve the quality of learned Bayesian networks. We previously presented Abstraction-Based Search (ABS), a technique for using attribute value hierarchies during Bayesian network learning to remove unimportant distinctions within the CPTs. In this paper, we introduce TCPT ABS (TABS), which integrates ABS with TCPT learning. Since expert-provided hierarchies may not be available, we provide a clustering technique for deriving hierarchies from data. We present empirical results for three real-world domains, finding that (1) combining TCPTs and ABS provides a significant increase in the quality of learned Bayesian networks (2) combining TCPTs and ABS provides a dramatic reduction in the number of parameters in the learned networks, and (3) data-derived hierarchies perform as well or better than expert-provided hierarchies.