Sparse self-representation subspace clustering algorithm for self-adaptive local structure embedding

Abstract

The invention discloses a sparse self-representation subspace clustering algorithm for self-adaptive local structure embedding. The invention belongs to the technical field of information. According to the invention, the optimal subspace and the most distinct clustering structure in the low-dimensional space can be identified at the same time, and the invention is superior to other two-stage subspace clustering methods; in addition, a nonlinear manifold regularizer is introduced, so that the learning trade-off between an original space and a subspace can be dynamically utilized; a local structure in an original space is encoded into a dictionary by a sparse self-representation method, and adaptive learning can be carried out in a clustering process. According to the invention, the non-square l2, 1-norm is adopted to minimize the residual error, and different from other methods based on the square l2-norm, the SSS can realize stable performance because the model based on the square l2,1-norm has robustness to abnormal values and noise; experimental results on an actual benchmark data set show that the method can provide more interpretable clustering results, and the performance ofthe method is superior to that of other alternative schemes.

Description

technical field [0001] The invention belongs to the field of information technology, specifically relates to a sparse self-representational subspace (SSS) clustering model with adaptive local structure embedding, specifically integrates dimensionality reduction and clustering into a unified model, and proposes a dynamic A new clustering method that makes use of the learning between the original space and the subspace and has stable performance. Background technique [0002] Clustering is a fundamental technique in data mining and has been applied in a wide range of fields, including image segmentation, recommender systems, and text mining. Among various clustering methods, methods based on matrix factorization (MF) have attracted extensive research. This method aims to approximate a given matrix by the product of two or more factor matrices. In this way, the K-means algorithm can be interpreted as a non-negative matrix factorization (NMF), so that the input data matrix can...

AI Technical Summary

Problems solved by technology

However, many existing graph-based methods use two-stage methods, which are sensitive to the quality of the adjacency graph, since redundant features in the input space hinder the construction of high-quality graphs.

Images