Feature sub-setting is through selection or elimination.
We look approximation algorithm:
- Feature evaluation criteria.
- Feature are ranked.
- Select top 'd' features.
1 and 2 could be independent of learning algorithm (unsupervised).
Example: Filtering algorithm (filters)/model.
Ex: Big data, independent of class label. Then it is called as unsupervised feature selection (classification).
1 and 2 could be dependent of learning algorithm (supervised).
Example: Wrapper algorithm (wrappers)/model.
Ex: Small data, class label/class information dependent. Then it is called as supervised feature selection (clustering).
