Hi Marcelo, 

Generally, the decision boundary of OC-SVMs is quite sensitive to outliers and your parameters, which is why you might want to bootstrap an ensemble. But bootstrapping OC-SVM models naively doesn't tend to perform wonderfully either, because most bootstrap samples will be relatively stable for each bootstrap stable. Certainly it won't perform optimally, but it won't perform terribly. 

You might want to consider weighting each bootstrap sample. The attached paper details an approach based on kernel density estimation, and gives a recipe to calculate the weights. Empirically, it seems that this approach works well. 

Niall

Hi Niall,

thanks for your answer. I will try to reproduce the paper's rsults to get a better understanding of the method. Thanks again!

Marcelo

Dear Marcelo, 

One way to improve the classification performance of an SVM based classifier 

is to try different type of kernel functions to see the one that gives the best classification performance for your dataset.

For example, we the following kernel types:

1.      Linear

2.      Polynomial

3.      Radial Basis Function

4.      Pre-computed Kernel

Generally, each kernel type is designed for some set of specific problems. So, suggest you try various kernel functions for dataset and adopt the one that gives you the best performance.

I hope this information will be helpful for you.

Good luck.

Thanks for the answers. Just for clarification, I have a bunch of OC-SVMs or generaly SVMs and would need to combine them in order to get a classification label. The paper from Niall shows an importance sampling method to select data samples to reduce errors in classification which is also interesting. Thanks for further remarks/answers to all!