A method for k-means-like clustering of categorical data

Abstract

Despite recent efforts, the challenge in clustering categorical and mixed data in the context of big data still remains due to the lack of inherently meaningful measure of similarity between categorical objects and the high computational complexity of existing clustering techniques. While k-means method is well known for its efficiency in clustering large data sets, working only on numerical data prohibits it from being applied for clustering categorical data. In this paper, we aim to develop a novel extension of k-means method for clustering categorical data, making use of an information theoretic-based dissimilarity measure and a kernel-based method for representation of cluster means for categorical objects. Such an approach allows us to formulate the problem of clustering categorical data in the fashion similar to k-means clustering, while a kernel-based definition of centers also provides an interpretation of cluster means being consistent with the statistical interpretation of the cluster means for numerical data. In order to demonstrate the performance of the new clustering method, a series of experiments on real datasets from UCI Machine Learning Repository are conducted and the obtained results are compared with several previously developed algorithms for clustering categorical data.