An Improved PSBCM-ALH Algorithm For Training Large Scale Support Vector Regression

With the development of artificial intelligence technology,the study of data-driven machine learning algorithms has attracted an increasing interest from various research areas.As a classic algorithm in machine learning,Support Vector Regression(SVR)has played an important role in the regression prediction because of its good generalization performance and suitability for nonlinear problems.However,most of the current research on SVR focuses on its application to small and medium-sized problems.To solve the problem that SVR solves slowly on large-scale data sets,this paper proposes an improved PSBCM-ALH algorithm for training large-scale support vector regression model,which is based on Proximal Stochastic Block Coordination Minimization-Augmented Lagrangian Homotopy(PSBCM-ALH)algorithm proposed by Wang et al.Firstly,this paper proves that the improved PSBCM algorithm is globally convergent in the ?-SVR model by improving the updating criteria of the working set.Furthermore,this paper improves the strategy for updating the working set by utilizing the partition organization of matrix in the ?-SVR model.After decomposition,this paper applies ALH algorithm to solve the subproblem which is a strongly convex quadratic programming problem.Finally,for linear kernel functions,this paper proposes an adaptive ALH algorithm,which adaptively adjusts the algorithm call mode by the degree of decline of the objective function.The numerical experiment results show that the improved PSBCM-ALH algorithm has obvious advantages to the well-known LIBSVM in terms of speed,accuracy and robustness.Besides,the improved PSBCM-ALH algorithm is also used to predict the stock index and futures prices,including the introduction of technical indicators,standardized processing feature,the definition of model forecast evaluation and the analysis of prediction results.The results show that the prediction effect of this algorithm is better than that of LIBSVM solver.Compared with the LIBSVM solver,which takes hundreds of seconds or even thousands of seconds,this algorithm takes about 18-30 seconds to reduces the objective function to a smaller value.