I've got ~1M points of {-1, 1}-labeled data in ~100 dimensions, each dimension standardized to have minimal values of 0 and maximal values of 1. Can anyone recommend some quick-and-easy heuristics for setting the kernel bandwidth of an RBF SVM? I'm sure there must be some great heuristics out there based on typical between-point distances. Thanks, much!
EDIT: Googling around has found me Jaakola's Heuristic, based on the median distance between any two points of opposite class. Other recommendations appreciated!
You can refer the article A practical guide to support vector classification by Hsu et al which recommends the rigorous grid search method to determine the optimum kernel parameters. To know quickly the different methods that are in use you can follow the recently published paper Soft sensing of particle size in a grinding process: Application of support vector regression, fuzzy inference and adaptive neuro fuzzy inference techniques for online monitoring of cement fineness or Online monitoring and control of particle size in the grinding process using least square support vector regression and resilient back propagation neural network
Suggested heuristic: B. Caputo, K. Sim, F. Furesjo, A. Smola, Appearance-based object recognition using SVMs: which kernel should I use? Proc of NIPS workshop on Statitsical methods for computational experiments in visual processing and computer vision, Whistler, 2002.
If you are programming in R you can find a good implementation in the sigest function of kernlab package
e.g. sigmaRangeReduced
- Badges
- Science topic