What are the criterion of the given dataset that makes it suitable for L1 SVM instead of L2? I know that it is favorable to use large dimensional features with L1 SVM to utilize its implicit feature selection but in my case even with large dimensions like 20000, L1 SVM lacking compared to L2. What are your suggestions to get improved results from L1 SVM?
Hi.
In fact, you could use any norm for regularization. L0 is on the sparse end, Linf is on the other end. Each problem has an optimal norm power. Which it is depends on many factors, and it is hard (perhaps impossible) to tell in advance.
In my understanding, sparsity should work if you have some good features and some noisy features --- the noisy feature should get suppressed. Higher norms should work for cases when the features have similar information content.
If you want the best results, use cross-validation to select the regularization.
Regards.
Hi,
I would try to improve the results by choosing different kernels.
You should use cross-validation to select the best for your data.
Good luck!
- Badges
- Science topic
- Similar topics
- Data Mining
- Classification