Discretization of Continuous Variables in Bayesian Networks Based on Matrix Decomposition

Abstract

Discretization of continuous variables (DCV) which could directly affect the results of Bayesian network inference (BNI) has been an important issue in Bayesian network (BN). Some common methods of DCV by the equal interval, the equal frequency, etc. always result in data loss which would make the results of BNI inaccurate. In this paper, a method of DCV in BN based on matrix decomposition is presented. This method could discretize the value of continuous variable into more states with different probability rather than one state, so it's more scientific and accurate. This paper makes a BN with two nodes, height and weight of each person, as an example and the simulation result demonstrates that the proposed method of DCV based on matrix decomposition can achieve discretization without data loss and ensure the accuracy of BNI.