In order to effectively extract nonlinear features of data set, the paper proposed a new method, called Kernel Semi-supervised Locality Preserving Projection (KSSLPP). It redefined the between-class similarity and within-class similarity using rich labeled and unlabeled samples that contain valuable information, which was used to maximize the between-class separability and minimize the within-class separability in a high dimensional kernel space. The proposed method preserves the global and local structures of unlabeled samples in addition to separating labeled samples in different classes. Contrast experiments in the Olivetti face database and UCI database verify the effectiveness of the proposed algorithm.