A Parameter Free BBN Discriminant Function for Optimum Model Complexity versus Goodness of Data Fitting

Abstract

Bayesian Belief Network (BBN) is an appealing classification model for learning causal and noncausal dependencies among a set of query variables. It is a challenging task to learning BBN structure from observational data because of pool of large number of candidate network structures. In this study, we have addressed the issue of goodness of data fitting versus model complexity. While doing so, we have proposed discriminant function which is non-parametric, free of implicit assumptions but delivering better classification accuracy in structure learning. The contribution in this study is twofold, first contribution (discriminant function) is in BBN structure learning and second contribution is for Decision Stump classifier. While designing the novel discriminant function, we analyzed the underlying relationship between the characteristics of data and accuracy of decision stump classifier. We introduced a meta characteristic measure AMfDS (herein known as Affinity Metric for Decision Stump) which is quite useful in prediction of classification accuracy of Decision Stump. AMfDS requires a single scan of the dataset.