Twin parametric-margin support vector machine (TPMSVM) obtains a significant performance. However, its decision function loses the sparsity, which causes the prediction speed to be much slow. In this brief, we present an improved TPMSVM, named centroid-based twin parametric-margin support vector machine (CTPSVM). The significant advantage of CTPSVM over twin support vector machine (TWSVM) and TPMSVM is that its decision hyperplane is sparse by optimizing simultaneously the projection values of the centroid points of two classes on its pair of nonparallel hyperplanes. In addition, a learning algorithm based on the clipping strategy is proposed to solve the optimization problems. Experimental results show the effectiveness of our method in speed, sparsity and accuracy, and therefore confirm further the above conclusion.