Parallel SVM Algorithms in Big Data Environments

Abstract

Nowadays, with the rapid development of information technology and data storage technology, the amount of data in various industries in society is growing rapidly, and the human society has stepped into the era of big data. How to efficiently process these massive data and extract potential value information from them has been a hot topic in recent years, and data mining technology has become the mainstream direction for people to explore this topic. Among the classification algorithms, Support VectorMachine (SVM) has become a popular research because of its good generalization ability, good ability to overcome dimensional disasters and nonlinear problem solving ability. However, traditional SVMs are only suitable for solving small sample problems, and their computational complexity grows exponentially when dealing with large data sets, while the emergence of parallelization techniques has solved this challenge. Therefore, it is of great value and significance to study parallelized SVMs. The proposed Support Vector Machine Algorithm using K-means Clustering and Whale Optimization Algorithm based on Map Reduce-–MR-KWSVM, firstly, proposes KF (K-means and Fisher) strategy to delete redundant data, and uses the data set after deleting redundant data to train SVM, which reduces the sensitivity of SVM to redundant data; secondly, The Improved Whale Optimization Algorithm based on Nonlinear convergence factor and Self-Adaptive Inertia Weight-IW-BNAW were proposed, and the optimal parameters of SVM are obtained by using “IW-BNAW” algorithm to improve the parameter optimization ability of support vector machine; finally, in the process of using Map Reduce to construct parallel SVM, Time Feedback Strategy (TFB) is proposed for load scheduling of reduce nodes to improve the parallel efficiency of the cluster and achieve high parallel SVM. Another algorithm proposed in this paper-–PKBBTSVM (Parallel Multi-Class SVM Algorithm Based on K-means and Balanced Binary Tree), firstly, according to the distribution characteristics of redundant data in multi-category data sets, a DKR strategy (Data Reduction Strategy Based on K-means and Sample Redundancy, DKR) is proposed, which speeds up the training speed of multi-classification SVM. Secondly, we propose the BSD strategy (Non-leaf node data partitioning strategy based on balanced binary tree and sample Divisibility) for building sub-classifiers of multi-category SVMs, and finally, the parallelization of multiclassification SVMs on Map Reduce is implemented.