Bayesian Network Learning with Abstraction Hierarchies and Context-Specific Independence

Context-specific independence representations, such as treestructured CPTs, reduce the number of parameters in Bayesian networks by capturing local independence relationships. We previously presented Abstraction-Based Search (ABS), a technique for using attribute value hierarchies during Bayesian network learning to remove unimportant dis-ion-Based Search (ABS), a technique for using attribute value hierarchies during Bayesian network learning to remove unimportant distinctions within the CPTs. Recently, we have recognized that the abstraction performed by ABS is complementary to that of TCPTs. In this paper, we introduce TCPT ABS (TABS), which integrates ABS with TCPT learning. Since expert-provided hierarchies may not be available, or may not provide the most useful distinctions, 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 dramatic reduction in the number of parameters in the learned networks, without loss of accuracy, and (2) data-derived hierarchies perform as well or better than expert-provided hierarchies.