Nvidia Interview Question | SVM
Question
Is it beneficial to perform dimensionality reduction before fitting an SVM? Why or why not?
in progress
1
Statistics
4 years
1 Answer
1036 views
Great Grand Master 0
Answer ( 1 )
Usage of simple normalization techniques such as feature scaling and mean normalization can often result in good accuracy rather than using PCA with SVM.
Eventhough PCA can help improve the discriminative power of classifiers, this doesn’t go well with SVMs since their kernel computation is not feature wise.
We could look at the distribution of Eigen values for the covariance matrix of our data, and see if they get very small. In general SVM’s are pretty robust in the cases where your data spans of the full feature dimension. The reason for this is that the SVM operates at the sample level (the kernel is computed between samples) and not at the feature level