The Research On Adaptive Scales Spectral Clustering Based On Nearest Neighbors Path

As a new clustering method, spectral clustering has become one of the most popularclustering algorithm in the last ten years, especially in the field of data mining and machinelearning. Compared to traditional clustering methods, spectral clustering can handle arbitrarydistribution data set, converge to global optimal and be easy to implement, so it is verysuitable for many practical problems. As spectral clustering algorithms are based on thesimilarity matrix, the similarity definition is very important for the performance of spectralclustering algorithm.In this paper the knowledge and methods of spectral clustering are firstly described,Then the similarity measure used in the classical spectral clustering, the Gaussian kernelfunction, are analyzed and researched in detail from the impact of the choice of the scaleparameter and the cluster consistency. The similarity measure of the classical spectralclustering can’t fully meet the cluster consistency and reflect the true approximaterelationship between two points due to not taking full account of local statistics and globalstructure. Based on these and combined with close neighbors thinking, the shared nearestneighbor weighted-based adaptive scale is firstly proposed, it takes full advantage ofstatistical information of close neighbors distribution; then k nearest neighbor path distance isproposed, it stresses the consistency of the global structure; Based on them, we proposednearest neighbor path-based adaptive scale similarity, the similarity well meets the clusterconsistency characteristics. Though introducing it into spectral clustering, adaptive scalesspectral clustering algorithm based on nearest neighbors path is got. Finally, a number ofexperiments on eight artificial data sets, five UCI data sets, and USPS data set are carried out.Experimental results show that the proposed algorithm achieves considerable improvementsover the typical spectral clustering algorithm and the Self-Tuning spectral clusteringalgorithm. It can reveal the relationship between the data points and find the real clusters.