Analysis of the Impact of data discretization in the bayesian network search space

Abstract

The structural learning of Bayesian Networks (BN) is a problem that has a search space with a super-exponential characteristic. Parameter choices affect structural learning and thus need to be studied. Considering that the most commonly used algorithms for structural learning of BNs require the use of quantized discrete data, quantization has to be studied. There are works that study the discretization and quantization of continuous variables for quantized discrete numeric variables and others that study the discretization and quantization of continuous variables for categorical variables. Some use information theory, others expectation-maximization and likelihood or other methods of discretization and quantization. However, all these works compare the final result of the structural learning or use a fixed structure and not how the quantization alters the search space of BNs. In this project we will conduct this study and create a thorough analysis of the implications of quantization in the search space for learning BNs. It is expected that our theoretical and experimental results will validate/invalidate results in the literature as to how the different types of quantization result in different qualities of learned BNs. (AU)