A Novel Discretization for Parameter Learning in Bayesian Network using Dynamic Programming

In AI and machine learning techniques such as decision trees and Bayesian networks, there is a growing need for converting continuous data into discrete form. Several approaches are available for discretization, however finding an appropriate and efficient discretization method is a challenging task. In this paper, we present an impurity based dynamic multi-interval discretization approach for learning Bayesian networks. This approach is based on the dependencies among nodes existing in the network and capable of discretizing the nodes in the network using the interdependencies with multiple class variables. The simulation based experiments conducted in this paper, suggest that the proposed discretization method produces good inference with very minimum information loss. The dynamic programming approach used in the method, further reduces the space and time complexity of the complete discretization process. Keywords-Bayesian Network; Discretization; Information Gain; Dynamic Programming