A Spectral Clustering-Based Dataset Structure Analysis and OutlierDetection Progress

摘要

To solve the problem that in real applications with spectral clustering algorithm, the number of clusters of a dataset is not given as a prior, a dataset structure analysis and outlier detection process is proposed in the paper. The proposed process is on the basis of spectral clustering algorithm and is consisted of four steps. The first step apply some algorithm, such as DBSCAN, which does not need the number of clusters as input to cluster the dataset to get an approximation of the number of clusters. The second step uses the approximation obtained from the first step to get the upper bound of the number of the clusters. The third step uses the upper bound obtained from the second step to get the optimal value of the number of the clusters, and output the optimal cluster result. The last step applies the LOF algorithm with the result from the third step to find the data objects with the largest probability to be outliers.