Efficient data classification by GPU-accelerated linear mean squared slack minimization

An efficient parallel implementation of the recently proposed Slackmin classification algorithm that minimizes the mean squared slack variables energy is proposed in this paper. The efficacy of the resulted scheme is demonstrated both in terms of accuracy and computation speed. The parallelization of the Slackmin algorithm is achieved in the framework of GPU programming. Based on this framework the “cuLSlackmin” algorithm for linear problems was implemented, by using the CUDA C/C++ programming model and proposed herein. The introduced parallel algorithm is making use of the advantages imposed by the GPU architecture and achieves high classification rates in a short computation time. A set of experiments with some UCI datasets have shown the high performance of the cuLSlackmin algorithm compared to the Slackmin, LIBSVM and GPULIBSVM algorithms. The high performance of cuLSlackmin algorithm makes it appropriate for big data classification problems.