Big data classification: A combined approach based on parallel and approx SVM

Abstract

This paper presents a combined solution for Big Data classification, by using one of the extended versions of the Support Vector Machines (SVM), known as the Parallel Support Vector Machines (PSVM). The main problem assumes that, once a PSVM model is obtained, a feature can be removed overtime, resulting in a decrease of the accuracy with the existing model. While Big Data is one of the interesting contexts, then training a new PSVM with the new data structure is time-consuming. The solution is to use an approach that approximates any SVM model based on the Radial Basis Function (RBF) kernel, and called the Approx SVM. In order to avert a new training step, this paper proposes to apply the Approx SVM in a parallel architecture. Despite that the Approx SVM was not purposely used to deal with large-scaled data set, the experimental results, which will be presented at the end of the article, are proofs that this approach is an appropriate choice for PSVM models.