Discretization of Numerical Attributes Preprocessing for Machine Learning

Page 2 of 46 Abstract The area of Knowledge discovery and Data mining is growing rapidly. A large number of methods is employed to mine knowledge. Several of the methods rely of discrete data. However, most datasets used in real application have attributes with continuously values. To make the data mining techniques useful for such datasets, discretization is performed as a preprocessing step of the data mining. In this paper we examine a few common methods for discretization and test these algorithms on common datasets. We also propose a method for reducing the number of intervals resulting from an orthogonal discretization by compromising the consistency level of a decision system. The algorithms have been evaluated using a rough set toolkit for data analysis. In Chapter 2, we introduce Knowledge Discovery. We also discuss the preprocessing step in the Knowledge Discovery pipeline, and introduce discretization in particular. Chapter 3 introduces some basic notions in Rough Set theory. In Chapter 4, we further discusses the discretization process, and investigates some common methods for discretization. In Chapter 5, we propose a two-step approach to discretization, using the Naive discretization algorithm introduces in 4.2, and a proposed algorithm to merge intervals. Empirical results from comparison of the different algorithms discussed in Chapter 4, and the proposed method from Chapter 5 can be found in Chapter 6. In Chapter 7, further work in the area of discretization is discussed. Appendix A contains some notes on the Rosetta[15] framework, and the implementation of the investigated algorithms in particular.