Abstract: As sparsity preserving projection algorithm is a globally unsupervised dimensionality reduction method, which turns out to be invalid to solve the lack of training samples or large differences in the same class, a locally sparsity persevering projection algorithm considering pairwise constraint is proposed. Sparse coefficients are obtained using neighbor samples in order to preserve local structure information. Then pairwise constraints supervision idea is taken for reference, applying sample labels to guide sparse reconstruction. Objective function is defined to catch the low-dimensional subspace which can best preserve the discriminant information contained in sparse coefficients. The effectiveness and feasibility of this proposed algorithm is verified by face recognition experiments from the perspective of recognition rate and runtime.

Key words: dimensionality reduction, locally sparsity preserving projection, pairwise constraint, face recognition

摘要： 稀疏保持投影算法是一种无监督的全局线性降维方法，无法应对训练样本不足及类内样本间差异过大的情况。针对该问题，提出一种结合成对约束机制的近邻稀疏保留投影算法。利用近邻样本求取稀疏系数以保留局部结构信息，引入成对约束监督的思想，利用样本类别指导稀疏重构过程，最后定义能最大限度保留稀疏系数中蕴含的类别信息的低维子空间。将该算法用于人脸识别，实验结果证明了算法在识别率以及运行时间上的有效性和可行性。

关键词: 降维, 近邻稀疏保留投影, 成对约束, 人脸识别