The TF terms depend on a single document, while the IDF terms depend on a collection of documents. If your test documents are already available when you are training your classifier, you can use both train and test sets to compute the IDF terms (this is a kind of semi-supervised learning). Otherwise, you can use the IDF values computed from the training set in the test phase. See also the Python example here:

http://scikit-learn.org/stable/tutorial/text_analytics/working_with_text_data.html#training-a-classifier

Hi,

 As I know, one of the main concerns in performance evaluation in machine learning techniques is that you should not use the training set for the test set too. In your case, if you use the TF/IDF or any feature vector values from your training set in your test set, your result will not be accurate enough. 

I hope it helps. 

Mostafa

Hi,

I guess it depends on the output of the training phase, if the output is a set of classification rules then you do not need to find the TF/IDF for the testing documents, you just need the terms vector to match with the generated rules then you calculate the accuracy..

This paper may help: https://www.researchgate.net/publication/232657035_Text_Classification_Based_on_Ant_Colony_Optimization

Best wishes...

Article Text Classification Based on Ant Colony Optimization

I agree with Qasem