Concerning that the samples are treated equally and the difference of samples is ignored in the conventional feature selection algorithms, as well as the learning model cannot effectively avoid the influence from the noise samples, an Unsupervised Feature Selection algorithm based on Self-Paced Learning (UFS-SPL) was proposed. Firstly, a sample subset containing important samples for training was selected automatically to construct the initial feature selection model, then more important samples were added gradually into the former model to improve its generalization ability, until a robust and generalized feature selection model was constructed or all samples were selected. Compared with Convex Semi-supervised multi-label Feature Selection (CSFS), Regularized Self-Representation (RSR) and Coupled Dictionary Learning method for unsupervised Feature Selection (CDLFS), the clustering accuracy, normalized mutual information and purity of UFS-SPL were increased by 12.06%, 10.54% and 10.5%, respectively. The experimental results show that UFS-SPL can effectively remove the effect of irrelevant information from original data sets.