Fast instance selection method for SVM training based on fuzzy distance metric

Support Vector Machine (SVM) is a well-known classification technique which has achieved excellent performance in many nonlinear and high dimensional pattern recognition fields. However, due to the high time complexity of training SVM model, it's difficult to implement it for large-scale data sets. One of the most promising solutions is to reduce the training data used for establishing the optimal classification hyperplane by means of selecting relevant support vectors which are the only factors affecting the classification rule. Thus, instance selection method is an efficient pre-processing technique to reduce the computational complexity and storage requirements of the learning process. In this manuscript, considering the geometry-distribution of data sets, we propose a Half Shell Extraction (HSE) algorithm which falls into the condensation category of instance selection methods. Moreover, fuzzy distance metric based on locality sensitive hash is employed to accelerate the instance selection process. Empirically, an experimental study involving various of data sets is carried out to compare the proposed algorithm with five competitive algorithms, and the results obtained show that the proposed algorithm consistently outperforms the other algorithms in terms of accuracy, reduction capability and runtime.