基于相关熵的快速聚类算法

Aiming at the issue of low efficiency and poor robustness in large-scale real-world data clustering, a fast correntropy-based clustering algorithm (FCC) is proposed. FCC is mainly divided into the following two steps: 1) Performing k-means on the original data to obtain rough labels to serve as the label matrix of the second step; 2) Adopting the original data and those anchors to construct the anchor graph and taking Laplacian matrix of the anchor graph as a graph constraint to explore the internal structure of original data so as to obtain the final categories of these samples. Meanwhile, the whole clustering process is carried out under the correntropy instead of the traditional Euclidean distance framework, which can effectively suppress the influence of the large amount of non-linear and non-Gaussian noise in the real-world data on the clustering robustness. To verify the performance of FCC, five state-of-the-art algorithms are developed as baselines to run with FCC on four large-scale real-world data sets. The results show that FCC can improve the clustering accuracy in most cases (by 8.58%, 6.86% and 1.86% respectively on WebKB, TDT2 and Cora) while greatly improving the clustering efficiency (by several or even dozens of times). Furthermore, to verify the robustness of FCC, varying degrees of random noise and Poisson noise are added to WebKB and Cora to obtain 8 noisy data sets, and all algorithms are run on these noisy data sets under the same conditions. Compared with the other baseline algorithms, FCC can maintain the optimal clustering robustness.