4
$\begingroup$

Given a classification task:

Training dataset "A" with labelled data of 10 classes.

Training dataset "B" with unlabelled data of 11 classes. Compared to "A", "B"contains one extra class, we can call it unknown class. Its size is also unknown. It represents the fact that there are certain "things" do exist in the world but it is not observed in our labelled training set.

Test dataset "C" contain 11 classes same as "B"

Any suggestion on how to handle this extra class during classification and predict "C" correctly?

Just want you know that this problem is related to the NIST iVector Challenge, the language identification task

Improve this question
$\endgroup$
2
  • $\begingroup$ Why is there no data in the training set with the "extra class"? Why not put "A", "B", and "C" together and randomly select "A", "B", and "C" again. Is the extra class a small sample size? $\endgroup$
    – Matt Reichenbach
    Apr 24 '15 at 12:26
  • $\begingroup$ HI, @MattReichenbach , the size of this extra class is unknown. It represents the fact that there are certain "things" that do exist but its not observed in our labelled training set. $\endgroup$
    – Steven Du
    Apr 25 '15 at 5:33
2
$\begingroup$

One way to deal with the problem of unknown classes is called open set recognition.

Walter J. Scheirer page http://www.wjscheirer.com/projects/openset-recognition/ has some papers of his on the subject.

Improve this answer
$\endgroup$

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.