Spectral Averagely-dense Clustering Based on Dynamic Shared Nearest Neighbors

Abstract

Spectral averagely-dense clustering is a clustering algorithm based on density, but it has the problem of being sensitive to the parameter ε. Aiming at the above problems, a spectral averagely-dense clustering based on dynamic shared nearest neighbors is put forward. Firstly, a similarity measures is constructed by combining self-tunning distance and shared nearest neighbors. Self-tunning distance can handle clusters of different density, and shared nearest neighbors can draw closer to the data in the same cluster and alienate the data in different clusters. Secondly, based on the sample distribution function, a method capable of self-adaptively determining the k-value of the shared nearest neighbors is proposed without setting the parameter k. Finally, the constructed similarity measure is used as the similarity measure of the fully connected graph. The ε-neighberhood graph of spectral averagely-dense clustering is replaced with the fully connected graph, which avoid setting the parameter ε. Through the experiments on artificial datasets and UCI datasets, the proposed algorithm is compared with the spectral averagelydense clustering and the standard spectral clustering. The experimental results show that the proposed algorithm not only avoids the problem of difficult selection of ε-neighberhood graph parameters, but also has better performance on the datasets.