A new approach to determine number of optimal clusters in partitioning clustering algorithms

A new approach to determine number of optimal clusters in partitioning clustering algorithms

Javad Arefnia1 Mohammad Amin Shayegan2

1) Islamic Azad University, Shiraz Branch Email:
2) Islamic Azad University, Shiraz Branch Email:

IOI Code : XAEC-CZHZC
Publication : The National Conference On New Approaches On Electrical And Computer Engineering(munce.ir)
Abstract :
    Data clustering is to find similar attributes among the massive amounts of data and their categorization into groups that each of these groups are called clusters. Since various factors such as noise and the number of dimensional data affect the outcome of different clustering algorithms, then these algorithms produce different results. Since the clustering quality and accuracy of the extracted clusters are very important, therefore, there are developed criteria for validating clustering operations. According to the information used to determine the quality of clustering, Clustering validation indicators are divided into internal and external categories. In this research, we have been evaluated three indices of internal standardization of clustering quality: Davies-Bouldin, Silhouette, and Gap. The effort of this research has been to suggest a new internal validation index, which would be better performed by using the partitional clustering algorithm K-Means and compared to other indices introduced on the standard dataset. The index introduced in this study is called CAS. The CAS s performance for detecting the number of clusters was 27.27% higher than the Davis-Bouldin index, and 36.56% more than the silhouette, and 54.54% better than the Gap index. Finally, it can be concluded that the CAS index with the most correct identification of the number of clusters has the most appropriate performance on the standard dataset than the other three standard indices.

Keywords : clustering validation clustering davies bouldin silhouette gap proposed cas index