Abstract:

Many feature transform methods have been proposed for the machine learning research area. They generally try to project the available data from the original feature space to a new feature space so that those data are more representative, or discriminative if they are intended to be assigned with some specific labels. General techniques mainly involve the Eigenvector or Spectral method, the optimization theories (Linear or Convex), the graph theories,and so on. It is generally (1) to construct a structure for the original data and their correlations, (2) to define an objective function to evaluate the purpose of the projection or the characteristics of the new space, (3) to apply optimization theories to optimize the objective function to get the solution to the problem. This paper gives two classical methods of locality preserving transformation. By analyzing their key points together with their deficiencies, we get a general view of the currently most critical problems.

Key words: feature transformation;PCA;LLE;LE;ISOMAP;NPE;LPP;LDE