Support Vector Machines on Large Data Sets: Simple Parallel Approaches
Support Vector Machines (SVMs) are well-known for their excellent performance in the field of statistical classification. Still, the high computational cost due to the cubic runtime complexity is problematic for larger data sets. To mitigate this, Graf et al. (Adv. Neural Inf. Process. Syst. 17:521–528, 2005) proposed the Cascade SVM. It is a simple, stepwise procedure, in which the SVM is iteratively trained on subsets of the original data set and support vectors of resulting models are combined to create new training sets. The general idea is to bound the size of all considered training sets and therefore obtain a significant speedup. Another relevant advantage is that this approach can easily be parallelized because a number of independent models have to be fitted during each stage of the cascade. Initial experiments show that even moderate parallelization can reduce the computation time considerably, with only minor loss in accuracy. We compare the Cascade SVM to the standard SVM and a simple parallel bagging method w.r.t. both classification accuracy and training time. We also introduce a new stepwise bagging approach that exploits parallelization in a better way than the Cascade SVM and contains an adaptive stopping-time to select the number of stages for improved accuracy.
- Border, A., Ertekin, S., Weston, J., & Bottou, L. (2005). Fast kernel classifiers with online and active learning. Journal of Machine Learning Research, 6, 1579–1619.Google Scholar
- Graf, H. P., Cosatto, E., Bottou, L., Durdanovic, I., & Vapnik, V. (2005). Parallel support vector machines: The cascade SVM. Advances in Neural Information Processing Systems, 17, 521–528.Google Scholar
- Karatzoglou, A., Smola, A., Hornik, K., & Zeileis, A. (2004). kernlab - An S4 package for kernel methods in R. Journal of Statistical Software, 11(9), 1–20.Google Scholar
- Schoelkopf, B., & Smola, A. J. (2002). Learning with kernels: Support vector machines, regularization, optimization, and beyond. Cambridge: MIT.Google Scholar