A novel sample reduction method for support vector regression ased on memory mode

Aims to solving the problem of training speed and memory taking during traditional support vector regression (SVR) training for large scale sample sets, a method based on memory mode is proposed in this paper, named memory mode support vector regression (MM-SVR). By simulating the memory law of human with a forgetting factor and considering the importance of data to simulating actual physical process, the offline history data is sampled by utilizing the timeliness of the observation data. The sampled data is taken as the training set, on which the model was gained by support vector regression. The simulation tests are carried out on several benchmark datasets. The results show that MM-SVR has advantages compared to RS-SVR and original SVR on training speed and robustness.