Improving the K2 Algorithm Using Association Rule Parameters

Abstract

A Bayesian network is an appropriate tool to work with the uncertainty that is typical of real-life applications. This chapter proposes a method for improving the K2 algorithm, reducing the set of allowable parents from which the algorithm selects actual parents and avoiding extra arc insertions. This new methodology uses data mining techniques, and in particular the computation of association rules parameters from a database of examples, to learn the structure of a Bayesian network. Association rules describe correlation of events, and are characterized by several parameters that can be used in structure learning. The chapter also presents an extension of K2 called K2-rules, which exploits a parameter normally defined in relation to association rules for learning Bayesian networks. K2-rules are compared with K2 and TPDA on the problems of learning four Bayesian networks. The results show that K2-rules improve both K2 and TPDA with respect to the quality of the learned network and K2 with respect to the execution time.