Appropriate Points Choosing for Subspace Learning over Image Classification

Dimension reduction techniques are very important, as high dimensional data is ubiquitous in many real world applications, especially in this era of big data. In this paper, we propose a novel supervised dimensionality reduction method, called appropriate points choosing based DAG-DNE (Apps-DAG-DNE). In Apps-DAG-DNE, we choose appropriate points to construct adjacency graphs, for example, it chooses nearest neighbors to construct inter-class graph, which can build a margin between samples if they belong to the different classes, and chooses farthest points to construct intraclass graph, which can establish relationships between remote samples if and only if they belong to the same class. Thus, Apps-DAG-DNE could find a good representation for original data. To investigate the performance of AppsDAG-DNE, we compare it with the state-of-the-art dimensionality reduction methods on Caltech-Leaves and Yale datasets. Extensive experimental demonstrates that the proposed Apps-DAG-DNE outperforms other dimensionality reduction methods and achieves state-of-the-art performance for image classification.