I-nice: A new approach for identifying the number of clusters and initial cluster centres

Abstract

This paper proposes I-nice, which is a new method for automatically identifying the number of clusters and selecting the initial cluster centres in data. The method mimics a human being in observing peaks of mountains in field observation. The clusters in a dataset are considered as the hills in a field terrain. The distribution of distances between the observation point and the objects is computed. The distance distribution is modelled by a set of Gamma mixture models (GMMs), which are solved with the expectation-maximization (EM) algorithm. The best-fitted model is selected with an Akaike information criterion variant (AICc). In the I-niceSO algorithm, the number of components in the model is taken as the number of clusters, and the objects in each component are analysed with the k-nearest-neighbour method to find the initial cluster centres. For complex data with many clusters, we propose the I-niceMO algorithm, which combines the results of multiple observation points. Experimental results show that the two algorithms significantly outperformed two state-of-the-art methods (Elbow and Silhouette) in identifying the correct number of clusters in data. The results also show that I-niceMO improved the clustering accuracy and efficiency of the k-means clustering process.