Adaptive Semisupervised Local Discriminant Analysis
| Autor: | Chih-Sheng Chang, 張志昇 |
|---|---|
| Rok vydání: | 2014 |
| Druh dokumentu: | 學位論文 ; thesis |
| Popis: | 102 Many references show that the hyper-dimensional data classification often suffer from the Hughes phenomenon. Feature extraction is an appropriate preprocessing for preventing the problem. Linear discriminant analysis (LDA) is a basic and often used supervised feature extraction method. However, the number of feature extraction by LDA depends on the number of classes. Recently, the semi-supervised local discriminant analysis (SELD) was proposed to overcome the drawback of LDA. SELD combines the scatter matrices of LDA and neighborhood preserving embedding (NPE) for taking into account the geometric property of neighbors. Hence, the between-class scatter matrix of SELD is non-singular, and more features can be extracted by SELD. However, the number of unlabeled samples is fixed in SELD. From the experiments on the Indian Pine data set, the number of unlabeled samples affects the classification performance a lot especially in the small sample size problem. That is the classification performance by applying SELD decreases from a specific number of unlabeled samples, when the number increases. In this paper, two parameters are considered to combine the scatter matrices of LDA and SELD, and an adaptive SELD is proposed to decrease the Influence about the number of unlabeled samples. In addition, the unlabeled sample were randomly selection by SELD and, hence, the local geometric property around the training samples cannot be preserved. So, the concept of the Voronoi diagram is applied to define the regions according to the training samples (labeled samples), and the unlabeled samples are chosen in the regions with respect to the training samples according to the nearest-neighbor rule. Experimental results on the education data set and two hyperspectral image data sets show that the proposed methods outperform SELD. |
| Databáze: | Networked Digital Library of Theses & Dissertations |
| Externí odkaz: |
|