Article ID | Journal | Published Year | Pages | File Type |
---|---|---|---|---|
532191 | Pattern Recognition | 2013 | 8 Pages |
••Graph model and parameter selection is time consuming and suffers from over fitting.••We approximate the intrinsic manifold by linear combination of several graphs.••The graph selection problem is replaced by the solution of multiple graph weights.••The factorization metrics and the graph weights are learned jointly and iteratively.
Non-negative matrix factorization (NMF) has been widely used as a data representation method based on components. To overcome the disadvantage of NMF in failing to consider the manifold structure of a data set, graph regularized NMF (GrNMF) has been proposed by Cai et al. by constructing an affinity graph and searching for a matrix factorization that respects graph structure. Selecting a graph model and its corresponding parameters is critical for this strategy. This process is usually carried out by cross-validation or discrete grid search, which are time consuming and prone to overfitting. In this paper, we propose a GrNMF, called MultiGrNMF, in which the intrinsic manifold is approximated by a linear combination of several graphs with different models and parameters inspired by ensemble manifold regularization. Factorization metrics and linear combination coefficients of graphs are determined simultaneously within a unified object function. They are alternately optimized in an iterative algorithm, thus resulting in a novel data representation algorithm. Extensive experiments on a protein subcellular localization task and an Alzheimer's disease diagnosis task demonstrate the effectiveness of the proposed algorithm.