SPID4.5: A Selective Pseudo Iterative Deletion Discretization Algorithm for Machine Learning, Uncertain Reasoning & Pattern Recognition

Many machine learning algorithms developed for classification, prediction and uncertain reasoning cannot handle continuous features. To use them on real world data sets, continuous attributes must be discretized into small number of distinct ranges. Also discretization provides an insight into critical values in continuous attributes. In this work (SPID4.5), an improvement of our previously published work SPID3, we iteratively compute pseudo deletion count at each boundary point of all continuous attributes and accept threshold points which reduce noise in the database most. The successive reduction of noise in the database results in a better discretization. Our extensive empirical experiments with real world datasets show that the state of the art algorithms for learning like CN2, C4.5, RISE and Naïve-Bayes give improvements in performance with SPID4.5 discretization and comparison with other supervised learning methods like MDLP and Chi-Merge show that SPID 4.5 achieves a highly competitive performance.