adrinjalali / network-classifier Goto Github PK

Right now LinearSVC is used to select features at each stage, and also used for prediction.

My tests showed that even though LinearSVC is good for selecting features, but using an l2 regularized SVM instead of an l1 regularized SVM, gives better performance if it's trained on the selected features given by the l1 regularized one.

Therefore features should be selected by LinearSVC, and then a normal l2-regularized has to be trained on the data using those features for predicting the output.

Implement the probabilistic model instead of current model for predicting the confidences of features and individual learners.