With the rapid development of technologies in metaverse, digital twins, virtual and augmented reality, three-dimensional (3D) point clouds have been widely applied to electric power, construction, advanced manufacturing, and other industries. As a result, how to reduce the redundancies of 3D point clouds data and how to effectively select useful point cloud features have played a critical role in the full use of massive point clouds data. Considering that most of the current feature selection methods pay little attention to specific instances, in this paper, we proposed a novel supervised feature selection method, named feature selection based on specific outliers optimization (FSSO). Specifically, in order to obtain accurate specific outliers (SOs), we first optimized the traditional mean center of class, and automatically defined the class majority. Then, we proposed the feature selection algorithm that could compute the intra-class relative deviation of SOs, and score features based on the deviations. Extensive experiments on 3D data clouds classification datasets (ModelNet40, IntrA, and ShapeNetCore), and on four high-dimensional handcrafted datasets show that the proposed FSSO can select discriminative features, and improve the classification accuracy.