Validating the effect of different discretization methods for redic K-prototype clustering algorithm

Abstract

�Abstract: The REDIC K-prototype clustering algorithm is designed for mixed datasets which selects the initial centroids significantly and it also removes the dependency on prior value for number of cluster (k) and influence parameter (λ). Data preprocessing on data set introduce empirical better performance for any data mining algorithm. In this paper the taxonomy is build by integrating the data preprocessing technique – discretization with REDIC K-Prototype clustering algorithm. This taxonomy validates the performance of the algorithm for four different dataset and three performance indices. The numerical attributes of dataset need to be discretized and converted to categorical attribute before the clustering. Here the four discretization techniques are considered Equal Width Binning, Equal Frequency Binning, Entropy Based Binning, and the special case of Equal Width Binning that is binary Binning Approach. The result of proposed algorithm are compared with the standard K-Mode and K-Prototype clustering for original dataset and discretized data set. From the performance analysis it is clear that for 70% cases the REDIC K-Prototype Clustering with different discretization method gives better performance in compare to standard algorithms.