Supervised And Semi-supervised Feature Combination Methods Based On Graph

In this paper, we study the feature combination of the supervised data andsemi-supervised data based on graphs, and propose a unified framework of supervisedfeature combination based on graphs (GF). And we take that Fisher discriminant analysis(FDA), principal component analysis(PCA), locality preserving projection (LPP),discriminant locality preserving projection (DLPP), marginal Fisher analysis (MFA), andmaximum margin criterion (MMC) add to GF, and obtain the corresponding methodswritten as GFDA, GPCA, GLPP, GDLPPP, GMFA and GMMC by means of differentselection of weighting matrixes. But the proposed framework meet with an oftensingularity problem of weighting matrixes, in order to deal with that problem, we makealgorithms for GF, that is regularized GF (RGF), GF based on null space (NGF), GF basedon singular value decomposition (GF/SVD), GF based on generalized singular valuedecomposition (GF/GSVD), and GF based on pseudo-inverse (PIGF), respectively.We propose semi-supervised GF (SGF) based on GF, and extend SGF with kerneltrick to obtain KSGF algorithm.A serise of experiments are made on AVIRIS remote sensing image in order toillustrate the efficient and effective of five feature combination methods. The resultsdemonstrate the effectiveness of the new feature combination. Particularly, GFDA is betterthan others, and KSGF is better than SGF.