What would be a good search range for Nu and Gamma Values in OneClassSVM if I want to do a grid search? I have ~17000 training samples of one class with each sample consisting of around 300 dimensions( I might do PCA/others and reduce dimensionality).
Hi,
A useful paper:
https://www.sciencedirect.com/science/article/abs/pii/S0950705114000379
and
https://github.com/benmack/oneClass/blob/master/notebooks/oneClassIntro.ipynb
here b. mack explained an occ classifier based on manual model selection
Ravi,
let us take each parameter one by one.
1. Nu is the parameter that controls the training errors (and the number of SVs). This parameter is always within the range (0,1].
2. Gamma parameter determines the influence of radius on the kernel. The range of this parameter depends on your data and application.
For example, in the article: Article One-class SVM for biometric authentication by keystroke dyna...
the values are chosen as:- Nu = [2-10 to 2-6] with steps 20.1
- Gamma = [2-40 to 2-13] with steps 21
As you already mentioned, grid search is the best way to get the optimal Nu-Gamma pair for classification. One suggestion is that you could use Hinge-Loss rather than RMSE to determine optimal parameters if you're performing binary detection [-1, +1]. However, if you aim to do a probabilistic detection, you could go for RMSE.
- Badges
- Science topic