基于随机子空间的局部鉴别投影*

针对高维数据容易对噪声敏感及容易造成维数灾难问题,文中提出基于随机子空间的局部鉴别投影算法(RSLDP).利用随机子空间方法对高维的原始数据进行特征选择,在生成的低维特征子空间构造近邻图,降低噪声影响.RSLDP通过最大化局部类间加权散度和最小化局部类内加权散度,同时最小化样本的总体局部散度,改进局部最大间距鉴别嵌入算法,较好刻画样本与其类间类内近邻中心点的关系,有利于鉴别特征的提取.在CMU PIE和AR这2个人脸数据库上的实验表明文中算法的有效性.

Local Discriminant Projection via Random Subspace

High dimensional data is sensitive to noise and the curse of dimensionality problem appears easily. A local discriminant projection algorithm based on random subspace(RSLDP) is proposed. The attributes of original high dimensional data are selected by random subspace method to generate a low dimensional subspace, and the nearest neighbor graphs are constructed in the low dimensional subspace. Thus, the influence of noise is reduced effectively. By RSLDP, the local inter-class weighted scatter is maximized, the local intra-class weighted scatter is minimized, and simultaneously the local scatter on data is minimized as well. Consequently, the performance of local maximal margin discriminant embedding (LMMDE) algorithm is improved.The relationship between the focusing point and its inter-class/intra-class nearest neighbor center point is well characterized by RSLDP. The effectiveness of the proposed algorithm is verified by the experiments on CMU PIE and AR face datasets.