Junction density based clustering algorithm for data with arbitrary shapes

Density-based clustering algorithms can deal with arbitrary shaped clusters in data. However, most of these algorithms face difficulties in handling large scale data, since they usually need to compute the distance between each pair of data points for density estimation. To alleviate this problem, we define a new type of density called junction density to measure the density of the junction region of two groups generated by K-means. Since the junction density is only computed for neighboring groups, the computation burden is small. Based on the junction density, we propose a new clustering method to merge the groups instead of directly clustering the data points. Specifically, it mines initial clusters in the groups then assigns the remaining groups to corresponding initial clusters. The experiments on several arbitrary shaped datasets demonstrate the efficiency and effectiveness of the proposed method.