Parallel Computing of Support Vector Machines: A Survey

The immense amount of data created by digitalization requires parallel computing for machine-learning methods. While there are many parallel implementations for support vector machines (SVMs), there is no clear suggestion for every application scenario. Many factor-including optimization algorithm, problem size and dimension, kernel function, parallel programming stack, and hardware architecture-impact the efficiency of implementations. It is up to the user to balance trade-offs, particularly between computation time and classification accuracy. In this survey, we review the state-of-the-art implementations of SVMs, their pros and cons, and suggest possible avenues for future research.