|
I have a binary classification problem. I have extracted 500 features from a set of 5000 samples using my domain knowledge. In other words, I have got hand crafted features. I wish to prove that these features actually are enough for performing classification and they make the 2 classes of samples separable. i.e. When the samples are represented with these features, there is exists a (reasonable) decision boundary. Please advice on how I can prove this. Is there any statistically appropriate way of measuring the significance of the set of features as a whole (NOT the significance of individual features)? |
|
You can use Mutual Information. Check the wiki page: http://en.wikipedia.org/wiki/Mutual_information If the features-subset that you select are really meaningful the mutual Information of this subset should be higher than the other. This means that you also have to calculate the mutual information of the other subset (not selected features). Also you can use P(C | X) where C is the class and X is the subset of features. You should be careful with the numerical result because they say which subset with the best but this does not say if this is the "optimal features separation". For example, a features 'f' that you select as 'good' could not be as good as you think. Best of luck =) |
