Adaptive Graph Fusion for Unsupervised Feature Selection

The massive high-dimensional data brings about great time complexity, high storage burden and poor generalization ability of learning models. Feature selection can alleviate curse of dimensionality by selecting a subset of features. Unsupervised feature selection is much challenging due to lack of label information. Most methods rely on spectral clustering to generate pseudo labels to guide feature selection in unsupervised setting. Graphs for spectral clustering can be constructed in different ways, e.g., kernel similarity, or self-representation. The construction of adjacency graphs could be affected by the parameters of kernel functions, the number of nearest neighbors or the size of the neighborhood. However, it is difficult to evaluate the effectiveness of different graphs in unsupervised feature selection. Most existing algorithms only select one graph by experience. In this paper, we propose a novel adaptive multi-graph fusion based unsupervised feature selection model (GFFS). The proposed model is free of graph selection and can combine the complementary information of different graphs. Experiments on benchmark datasets show that GFFS outperforms the state-of-the-art unsupervised feature selection algorithms.