Node Importance-Based Semi-supervised Nonnegative Matrix Factorization for Image Clustering

As a typical dimensionality reduction method, non-negative matrix factorization (NMF) is widely used in image clustering tasks. However, the traditional NMF is unsupervised and cannot fully utilize the label information. In this research, we propose a joint feature representation and node importance propagation framework called semi-supervised nonnegative matrix factorization based on node importance (NISNMF). Firstly, by considering the overlap degree of the shared neighbors of the labeled node, a new importance index is proposed based on local structure to increase the influence of important node. Secondly, NISNMF integrates semi-supervised non-negative matrix factorization and node importance propagation into a unified feature representation and information propagation framework to improve the discriminative ability of the model. Finally, an efficient alternating iteration method is designed and its convergence is proved. Experimental results in a large number of image clustering tasks verify the superiority of the algorithm. In addition, the accuracy, parameters and sensitivity of the algorithm are also experimentally investigated.