Improving the SLA Algorithm Using Association Rules

Abstract

A bayesian network is an appropriate tool for working with uncertainty and probability, that are typical of real-life applications. In literature we find different approaches for bayesian network learning. Some of them are based on search and score methodology and the others follow an information theory based approach. One of the most known algorithm for learning bayesian network is the SLA algorithm. This algorithm constructs a bayesian network by analyzing conditional independence relationships among nodes. The SLA algorithm has three phases: drafting, thickening and thinning. In this work, we propose an alternative method for performing the drafting phase. This new methodology uses data mining techniques, and in particular the computation of a number of parameters usually defined in relation to association rules, in order to learn an initial structure of a bayesian network. In this paper, we present the BNL-rules algorithm (Bayesian Network Learner with association rules) that exploits a number of association rules parameters to infer the structure of a bayesian network. We will also present the comparisons between SLA and BNL-rules algorithms on learning four bayesian networks.