Concept Factorization with Optimal Graph Learning for Data Representation

摘要

In recent years, concept factorization methods become a popular data representation technique in many real applications. However, conventional concept factorization methods cannot capture the intrinsic geometric structure embedded in data using the fixed nearest neighbor graph. To overcome this problem, we propose a novel method, called Concept Factorization with Optimal Graph Learning (CF_OGL), for data representation. In CF_OGL, a novel rank constraint is imposed on the Laplacian matrix of the initial graph model, which encourages the learned graph with exactly c connected components for the data with c clusters. Then the learned optimal graph regularizer is integrated into the model of concept factorization. Therefore, this learned structure is benefit to the clustering analysis. In addition, we develop an efficient and effective iterative optimization algorithm to solve our proposed model. Extensive experimental results on three benchmark datasets have demonstrated that our proposed method can effectively improve the performance of clustering.