Consider we have a binary SVM classifier, classifying event A and event B and we want to add another class (or more) say event C, to our SVM. Can we add C without using the training data of event A and B. Generally SVM is binary classifier so to use for more than one class we can use "1-against-the rest" or "1-against-1." For either of this method can I use only data from training of event C and the earlier SVM parameters of A and B, without data of A and B for training.

In my opinion we cannot train SVM with learned parameters alone we need data from A and B as well. The question is: is it true? I am looking for an article or mathematical reason?


Assuming the following from your question:

  1. Data population has only 3 classes (A,B and C)
  2. You have tagged data for two classes only (say A and B).

What you can do for the third class (assuming you have no tagged data for A and B) is outlier detection. You can train One-Class SVMs on A+B data and if your model says its an outlier, its class C. You still have to run a standard classifier for detection of class A and class B.

  • $\begingroup$ The question "with A and B" training data.. $\endgroup$ – Creator Aug 23 '18 at 2:20
  • $\begingroup$ In point 2, Tagged data = Training data, and in point 1 data population = Train Data + Test data. Hope it clarifies. $\endgroup$ – silent_dev Aug 24 '18 at 15:13
  • $\begingroup$ Thanks for your replies,but I am afraid I am not looking for this answer. What happens, if we have one more class say D? There is no way (my opinion) we can add new classes one after another without retraining with all data. I want a mathematical proof or article to show or explain why this is true. $\endgroup$ – Creator Aug 27 '18 at 2:08
  • $\begingroup$ If you write an answer to the question stats.stackexchange.com/questions/364445/… I would love to accept the answer. $\endgroup$ – Creator Aug 29 '18 at 1:00

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.