A Robust Self-Organizing Approach to Effectively Clustering Incomplete Data

Abstract

In the real world, incomplete data are often encountered and located anywhere in a data set. Such incomplete data make a data clustering task more challenging. It's common practice to eliminate incomplete data from the input data set. If there are a large number of missing values, ignoring them may lead to the data insufficiency and ineffectiveness of the data clustering task. Hence, incomplete data clustering has been considered in many research works with many different approaches based on the well-known existing clustering algorithms such as k-means, fuzzy c-means, the self-organizing map (SOM), mean shift, etc. However, few of them have examined both effectiveness and robustness of the incomplete data clustering algorithms. Some of them are not practical due to a lot of parameters in hybrid approaches and/or cannot handle incomplete data which appear in any object at any dimension. In contrast, this paper aims at a SOM-based incomplete data clustering algorithm, iS nps, which is a robust and effective solution to clustering incomplete data in a simple but practical approach. Is nps can do clustering on incomplete data as well as estimate incomplete data using the nearest prototype strategy in an iterative manner. As compared to several different existing approaches, our proposed algorithm can produce the clusters of good quality and a better approximation of incomplete data via the experiments on benchmark data sets.