RDMN: A Relative Density Measure Based on MST Neighborhood for Clustering Multi-Scale Datasets

Density based clustering techniques discover the intrinsic clusters by separating the regions present in the dataset as high- and low-density regions based on their neighborhood information. They are popular and effective because they identify the clusters of arbitrary shapes and automatically detect the number of clusters. However, the distribution patterns of clusters are natural and complex in the datasets generated by different applications. Most of the existing density based clustering algorithms are not suitable to identify the clusters of complex pattern with large variation in density because they use fixed global parameters to compute the density of data points. Minimum spanning tree (MST) of a complete graph easily captures the intrinsic neighborhood information of different characteristic datasets without any user defined parameters. We propose a new Relative Density measure based on MST Neighborhood graph (RDMN) to compute the density of data points. Based on this new density measure, we propose a clustering technique to identify the clusters of complex patterns with varying density. The MST neighborhood graph is partitioned into dense regions based on the density level of data points to retain the shape of clusters. Finally, these regions are merged into actual clusters using MST based clustering technique. To the best of our knowledge, the proposed RDMN is the first MST based density measure for capturing the intrinsic neighborhood without any user defined parameter. Experimental results on synthetic and real datasets demonstrate that the proposed algorithm outperforms other popular clustering techniques in terms of cluster quality, accuracy, and robustness against noise and detecting the outliers.

Automated summary

Density-based clustering techniques are popular because of their ability to identify clusters of arbitrary shapes and automatically detect the number of clusters. However, these techniques are not suitable for clusters with large density variation. We propose a new density measure based on minimum spanning tree (MST) and a clustering technique to address this issue.