FS Methods | ||
---|---|---|
Advantages | Disadvantages | |
Filter | They are easily scalable to very high-dimensional data sets. | They do not interact with the classification algorithm. |
They are computationally fast and simple. | Most of this methods are univariate, this is, they consider features independently or only with regard to the target feature, thereby ignoring feature dependencies. | |
They are independent of the classification algorithm used in the further model construction. | ||
Wrapper | They include the interaction between feature subset search and the classification algorithm that is “wrapped”. | They have a higher risk of overfitting, depending on how exhaustive is the feature subset search. |
They take into account feature dependencies. | They are very computationally intensive, especially if the “wrapped” classifier has a high computational cost. | |
Embedded | They include the interaction between feature subset search and the final classification model constructed. | They depend on the specific learning method of the final model constructed. |
They take into account feature dependencies. | ||
They are computationally faster than wrapper methods. |