A distributed SVM method based on the iterative MapReduce

Abstract

Linear classification is useful in many applications, but training large-scale data remains an important research issue. Recent advances in linear classification have shown that distributed methods can be efficient in improving the training time. However, for most of the existing training methods,based on MPI or Hadoop, the communication between nodes is the bottleneck. To shorten the communication between nodes, we propose and analyze a method for distributed support vector machine and implement it on an iterative MapReduce framework. Through our distributed method, the local SVMs are generic and can make use of the state-of-the-art SVM solvers. Unlike previous attempts to parallelize SVMs the algorithm does not make assumptions on the density of the support vectors, i.e., the efficiency of the algorithm holds also for the “difficult” cases where the number of support vectors is very high. The performance of the our method is evaluated in an experimental environment. By partitioning the training dataset into smaller subsets and optimizing the partitioned subsets across a cluster of computers, we reduce the training time significantly while maintaining a high level of accuracy in both binary and multiclass classifications.