Experimental Evaluation of Three Partition Selection Criteria for Decision Table Decomposition

Decision table decomposition is a machine learning approach that decomposes a given decision table into an equivalent hierarchy of decision tables. The approach aims to discover decision tables that are overall less complex than the initial one, potentially easier to interpret, and introduce new and meaningful intermediate concepts. Since an exhaustive search for an optimal hierarchy of decision tables is prohibitively complex, the decomposition uses a suboptimal iterative algorithm that requires the so-called partition selection criterion to decide among possible candidates for decomposition. This article introduces two such criteria and experimentally compares their performance with a criterion originally used for the decomposition of Boolean functions. The experiments highlight the differences between the criteria, but also show that in all three cases the decomposition may discover meaningful intermediate concepts and relatively compact decision tables.