PROJECT TITLE :

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

ABSTRACT:

Techniques for discovering intrinsic clusters that are based on density do so by classifying the regions that are present in the dataset into high-density and low-density regions according to the information about their surrounding areas. Because they can determine the clusters of arbitrary shapes and automatically count how many clusters there are, they have gained a lot of popularity and are very useful. On the other hand, the patterns of cluster distribution are both natural and complex in the datasets that are generated by the various applications. Because they use fixed global parameters to compute the density of data points, the vast majority of the currently available density-based clustering algorithms are not suited to identify the clusters of complex patterns with large variations in density. This is because of how the algorithms work. The minimum spanning tree, also known as MST, of a complete graph is able to easily capture the intrinsic neighborhood information of various characteristic datasets without the need for any user-defined parameters. In order to compute the density of data points, we propose a new Relative Density measure that is based on the MST Neighborhood graph (RDMN). We propose a clustering technique to identify the clusters of complex patterns with varying densities, and it is based on this new density measure that we have developed. In order to maintain the cluster-like structure of the MST neighborhood graph, it is divided into dense regions that are proportional to the number of data points in each region. In the end, actual clusters are formed by combining these regions into them using the MST-based clustering technique. The proposed RDMN is the first MST-based density measure for capturing the intrinsic neighborhood without any user-defined parameters, to the best of our knowledge. The proposed algorithm outperforms other popular clustering techniques according to experimental results on both synthetic and real datasets. These results demonstrate that the proposed algorithm is superior in terms of cluster quality, accuracy, and robustness against noise in addition to detecting outliers.