부산대학교 도서관

Various graph-based unsupervised feature selection (UFS) methods have recently been developed to exploit the hidden structure information in data, achieving remarkable results. However, there are still some challenges: (1) most of them construct a suboptimal similarity graph to capture the global or local structure individually; (2) they ignore the high-order relations (i.e., the relation between the neighbor structure among samples). To tackle these challenges, Dynamic Optimal Graph Learning for Unsupervised Feature Selection (DOGL) is proposed to learn an optimal graph, which can simultaneously capture the global and local structure in embedding subspace. In particular, we project the original data into a low-dimensional embedding subspace to remove the corrupted information (i.e., noisy features, redundant features, and outliers) in the data. Then, a global similarity graph is learned based on data reconstruction in the embedding subspace. Meanwhile, we learn a mixed-order similarity graph from the predefined local similarity graph to exploit the high-order neighborhood information. Furthermore, we incorporate the two processes into a united framework, which can learn an optimal graph to capture the global and local structure while eliminating the influence of corrupted information. Last but not least, we design an iterative algorithm to solve the objective function of DOGL. Extensive experiments on six datasets demonstrate the effectiveness and superiority of DOGL.