Abstract: Dimensionality reduction without losing intrinsic information on original data is an important technique for succeeding tasks such as classification. A novel local within-class features preservation kernel fisher discriminant algorithm was proposed after deeply analyzing the relationship between kernel fisher discriminant and kernel local fisher projection. I}he new method keeps the ability of KFD's global projection and solves the over-fitting of KI_FDA's local preservation problem, which can work well on overlapped or multimodal labeled data. I}he training algorithm is improved for resolving out of-memory problem when applied in large sample situation. The simulation and speaker identificanon application show that the proposed algorithm has more adaptability as well as advanced recognition rate and speed.

Key words: Fisher discriminant analysis, Local preservation projection, Speaker identificaiton, Kernel trick, Dimensionality reduction