7.7: SVM
Support vector machines learn very
well from high-dimensional data. When we have more features, SVM uses the extra
information from additional features and tries to build a better model. This is
true even when multicollinearity exists. Hence, feature selection doesn't help
mostly in improving performance for SVM.
If overfitting is of concern, we can
instead select a suitable value of SVM parameter C through experimentation. It
intuitively acts like ridge regression. In that it doesn't remove variables. It
only changes the level of impact a specific feature has. This is best suited
for situations when the feature set has many binary features, which increases
degrees of freedom.