Analysis and Implementation of Large-scale Linear RankSVM in Distributed Environments

Linear rankSVM is a useful method to quickly produce a baseline model for learning to rank. Although its parallelization has been investigated and implemented on GPU, it may not handle large-scale data sets. In this thesis, we propose a distributed trust region Newton method for training L2-loss linear rankSVM with two kinds of parallelizations. We carefully discuss the techniques for reducing the communication cost and speeding up the computation, and compare both kinds of parallelizations on dense and sparse data sets. Experiments show that our distributed methods are much faster than the single machine method on two kinds of data sets: one with its number of instances much larger than its number of features, and the other is the opposite.