GLOBALITY-LOCALITY BASED ADAPTIVE GRAPH LEARNING FOR DIMENSIONALITY REDUCTION

When the traditional dimensionality reduction method based on graph learning processes real data, due to the noise and redundant information contained in the data, the information expression of the graph composed of the original data is inaccurate, which will affect the classification effect. In order to solve this problem, a globality-locality based adaptive graph learning for dimensionality reduction method is proposed. We preserved the local structure of the data by adding graph constraints to the reconstruction residuals of the data. We reconstructed low-dimensional data and minimized reconstruction errors to preserve the global structure of the data while learning the projection matrix. In addition, the similar matrix in the graph learning was used as a low-dimensional reconstruction matrix, so that the learned graph contained both the local structure and the global structure of the data, so as to more accurately express the relationship between the data. Experimental results on three databases show that this method can obtain better classification results.