A Comparative Study of Several Cluster Number Selection Criteria

Abstract

Abstract. The selection of the number of clusters is an important and challenging issue in cluster analysis. In this paper we perform an experimental comparison of several criteria for determining the number of clusters based on Gaussian mixture model. The criteria that we consider include Akaike's information criterion (AIC), the consistent Akaike's information criterion (CAIC), the minimum description length (MDL) criterion which formally coincides with the Bayesian inference criterion (BIC), and two model selection methods driven from Bayesian YingYang (BYY) harmony learning: harmony empirical learning criterion (BYY-HEC) and harmony data smoothing criterion (BYY-HDS). We investigate these methods on synthetic data sets of different sample size and the iris data set. The results of experiments illustrate that BYY-HDS has the best overall success rate and obviously outperforms other methods for small sample size. CAIC and MDL tend to underestimate the number of clusters, while AIC and BYY-HEC tend to overestimate the number of clusters especially in the case of small sample size.

A comprehensive survey of traditional, merge-split and evolutionary approaches proposed for determination of cluster number

“…Hu and Xu [91] presented a comparative study of the probabilistic criteria, including Akaike's information criterion (AIC) [92], consistent Akaike's information criterion (CAIC) [93], minimum description length (MDL) [94], Bayesian Ying-Yang (BYY) harmony empirical learning criterion (BYY-HEC) [72] and BYY harmony data smoothing learning criterion (BYY-HDS) [95] using Gaussian mixture based clustering. The results indicated that BYY-HDS was superior to the other criteria.…”

Enhanced Dark Block Extraction Method Performed Automatically to Determine the Number of Clusters in Unlabeled Data Sets

“…Guo [21] proposed a cluster number choice method for a small set of samples using a Bayesian Ying-Yang (BYY) model. Comparative studies such as [17] and [22] provided experimental comparisons of many criteria such as Akaike's Information Criterion (AIC), Minimum Description Length (MDL), and (BYY) for determining the number of clusters based on a Gaussian mixture model. A variety of statistical techniques for tendency assessment are discussed in the work of Jain and Dubes [3].…”

Time Series Clustering Via RPCL Network Ensemble With Different Representations

“…Our visualization analysis on five datasets reveals that there are extremely complex cluster structures, i.e., the datasets contain a large number of classes, unbalanced cluster membership, low intracluster dissimilarity, high intercluster similarity, high dimensionality, and arbitrary cluster shapes, which results in failures for existing clustering algorithms including ours. Moreover, our empirical studies along with others [21], [32], [33], [37]- [40] indicate that our RPCL network ensemble on four representations described in Section II performs well for only the data with balanced and a small number of clusters structures, which is caused by the limitation of the coarse representations and RPCL network.…”