DCSNE: Density-based Clustering using Graph Shared Neighbors and Entropy

Abstract

Density-based clustering techniques identify arbitrary shaped clusters in the presence of outliers by capturing the intrinsic distribution of data and separating high and low-density regions based on the neighborhood information. They use global parameters to compute the density distribution of data points, the optimization of which is quite a challenging and time intensive task. The similarity graphs constructed from the data points can easily capture the topology of the density regions without using any user-defined parameters. Moreover, the concept of entropy can be useful to capture the randomness and disorderliness present in highly complex data distributions. Our proposed algorithm makes use of entropy in conjunction with the graph local neighborhood information to compute density distribution of data points. Then, actual clusters are identified using the density distribution and the shared neighbor information between the regions. The experimental results show that the proposed technique outperforms other comparable methods in terms of cluster quality in the presence of noise on diversified gene expression and real datasets.