Normalized Cheeger Cut With Neighborhood Rough Approximation

Abstract

A graph-based collecting study recently attracted significant attention from the scientific research community. Normalized Cheeger cut is a balanced graph partition criterion and a generalized version of normalized graph cut. A stress-free resolution of the normalized Cheeger cut can be obtained by employing the eigenvectors of curve p-Laplacian. However, it is highly sensitive for the original Cheeger cut to collect the interference of noise and disrelated properties. Thus, the performance of the Cheeger cut decreases when high-dimensional data are grouped. To decrease the negative influence of outliers and superfluous properties of collecting, we design an efficient attribute decrease method which is based on neighborhood rough approximation. This design aims to improve the collection of the Cheeger cut. The suggested algorithm introduces information entropy to the neighborhood rough sets in order to measure the importance of attributes. This algorithm reserves the most valuable features and removes the redundant features while retaining the maximum category information of raw data. We then build the p-Laplacian array with the optimized attribute sets and obtain the collecting consequences via the eigen-subspace decomposition of graph p-Laplacian. The cogency of the proposed algorithm is established in various standard data collections. Experimentations demonstrated that our method enjoys sturdy robustness to noise or disrelated feature information in high-dimensional figures.