A Clustering Method Based on Neighborhood Chain

One major difficulty in clustering is to discover the underlying structure of the input data space under complex situations without much prior information. This problem consists of two aspects, i.e., to determine the target clusters number based on the input data itself, and to properly assign the data into each cluster. Various methods have attempted to address the problem and achieved success in many applications. However, their performance is not satisfactory when the tasks need to be done under complex situations including (but not limited to) the clusters in non-spherical shapes and different densities. This paper proposes a clustering algorithm based on the neighborhood chain, a closeness measure different from the geometric distances used in most of the existing clustering methods. The proposed algorithm performs effectively for clusters with arbitrary shapes and different densities, and is able to help users determine the proper clusters number in an intuitive way. Experiments and comparisons are implemented and the results have demonstrated the effectiveness of the new clustering method.