0
$\begingroup$

I have a soft classfication problem, i.e., the correct label for a certain instance is not just one class with 100% probability, but rather bunch of classes with probabilities that sum up to one.

What I know as apriori information (I know it because I understand the underlying physical phenomenon) that the classes are close together, like a cluster. E.g, class one and five and eight always come together cause they're adjacent (PS: I can create adjacency matrix if required).

What do I want?

I want some way to tell the NN about this fact. Currently it is vanilla classfication neural net. Any suggestions or readings or guidannce are appreciated.

Improve this question
$\endgroup$

1 Answer 1

Reset to default
1
$\begingroup$

Why not create another set of classes which corresponds to the groups you want? This works if you care more about groups. Later on you can subsequently classify within groups. Another idea would be to devise a loss function which penalize less within group misclassification and more outside group classification but this involves going into implementation details of your NN. As an idea it would involve writing something like softmax multiplied with misclassification cost matrix

Improve this answer
$\endgroup$
1
  • $\begingroup$ Thanks, that's exactly what I did. I added one more term to my loss in which I penalized for non-adjacent classes. I was thinking of something more fancy like probabilisitc approach, but this is really simple and effective. $\endgroup$
    – Alex Deft
    Commented Jul 20, 2019 at 21:59

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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