A Heuristic Initialization-Independent Spectral Clustering

As an unsupervised learning method, spectral clustering is widely used in machine learning. Spectral clustering algorithms apply spectral graph theory to solve the clustering of non-convex sphere of sample spaces, so that they can be converged to global optimal solution. After the essence of initialization sensitivity in spectral clustering is analyzed, the prim algorithm is used to select the k initial cluster centers, and a threshold parameter is set to solve the problem that two data objects in the same class are selected as the cluster centers, which leads to get the wrong cluster results and increases the iteration times. Based on above, a heuristic initialization-independent spectral clustering algorithm is proposed. Compared with the traditional spectral clustering algorithm and the initialization-sensitive k-means algorithm, the experiments show that the suggested algorithm not only obtains the stable cluster centers and cluster results, but also effectively decreases the iteration times, accordingly improves the clustering performance, which validates the feasibility and effectiveness of the suggested algorithm.