Corpus Characteristics-Based Method to Centroids Number Determination for Clustering Text Documents

Clustering is fundamental for categorizing and exploring information, particularly in written texts. Traditional clustering algorithms such as K-Means generate clusters in which each document is assigned to a cluster based on a centroid or representative of the cluster. Nevertheless, it is common to find situations where a single centroid is not enough to represent a cluster. To solve this problem, some variants of the K-Means algorithm have been introduced by considering more than one centroid per cluster. However, determining the number of centroids per cluster is a challenge; the user must employ trial and error to determine a good value, which is time-consuming. This paper proposes a solution to this problem for Improved-FPAC (Fast Partitional Clustering Algorithm, one of the most recent text document cluster algorithms) by introducing a method to compute the parameter l (number of centroids per cluster) based on the characteristics of the corpus. Based on our experiments on different public standard data sets, the method proposed in this paper, allows Improved-FPAC to obtain better clustering quality than the default value suggested by Improved-FPAC’s authors.