DOE-AND-SCA: A Novel SCA Based on DNN With Optimal Eigenvectors and Automatic Cluster Number Determination

Abstract

Spectral clustering algorithm (SCA) is one of the widely used clustering algorithms (CAs), which is proved to be efficient in many applications including unsupervised image identification and gene prediction. However, most SCAs are confronted with several problems: 1) It is difficult for SCAs to handle multi-scale data sets; 2) It is difficult to set cluster number in advance for various applications; 3) It is also difficult to choose the most appropriate eigenvectors to reflect the data distribution; and 4) Moreover, SCAs are sensitive to the parameters. To handle these problems, we propose a novel SCA based on dynamic nearest-neighbors (DNN) with optimal eigenvector and automatic cluster number determination, namely DOE-AND-SCA. There into, first, we design a novel similarity function based on DNN for multi-scale data, making the similarity metric more accurate; Second, the cluster number is automatically determined, and the cluster centers are also automatically determined by normal fitting, based on the density and minimum distance distribution of the data points; Third, the optimal eigenvectors are selected on the basis of global and local features of the data set for more accurate data distribution reflection; Fourth, two main parameters, including the optimal density difference threshold and the number of intervals, are self-adaptive. The efficiency of DOE-AND-SCA is testified on abundant of simulation data sets, by comparing with other outstanding algorithms. And finally, DOE-AND-SCA is also applied to image recognition problems.