Detecting cluster numbers based on density changes using density-index enhanced Scale-invariant density-based clustering initialization algorithm

Despite high accuracy, K-means relies mainly on the determination of the suitable number of clusters. To cope with, it is hypothesized that in a dataset region with high density tends to be a cluster. The present study is based on Scaleinvariant density-based clustering initialization, in which a cluster numbers is derived from density change analysis or density distribution analysis. However, the density calculation under this approach is based on the number and volume of data, which may result in inaccuracy for cluster detection. Thus, the objective of this study was to improve the performance of Scaleinvariant density-based clustering initialization to detect the appropriate cluster numbers and initial cluster centers. We proposed a density calculation based on data distance. The density value obtained from the calculation was used as a condition of data division and data merging for cluster detection. According to the experiment, compared to the Scale invariant density-based clustering initialization, the proposed method could detect the cluster numbers and initial cluster centers more equal or closer to the actual number of clusters. In addition, the level of accuracy in clustering was higher than its counterpart.