0
$\begingroup$

I would like to classify my data set into several classes which have a size relation (between classes) using supervised learning. For example let's say I'd like to classify subjects in order to predict their shirts' size.

My features would be stuff measured like:

 BMI, height, weight, etc... 

The classes would be:

 XXS,XS,S,M,L,XL,XXL,XXXL

The classes have a relationship between them i.e.: If my true value is M and the algorithm predicts S is better than predicting XXL. Which supervised classification algorithm would you use in such a case?

$\endgroup$
0
$\begingroup$

It seems that the features linearly reflect the prediction results.

Thus, a fully connected neural network with one softmax output layer and no hidden layers would be predictive enough of the classes, and this is a good start. If not performing well, add a hidden layer before the softmax output layer, and the performance would be good enough.

$\endgroup$
  • $\begingroup$ It's not really what I've asked. I'm not looking what is the relationship between the features and the classes (this could be a non linear function). I'm really looking for an algorithm which can separate a case where the false prediction is close to the real class and cases where a false prediction is far from the real class. One more example: let's say the real class is XL. I would like an optimization function which would favor a L prediction over XS prediction. Both predictions are FALSE but because the classes are related we can say the L prediction is better. $\endgroup$ – AR_ Jul 25 '17 at 11:34
  • $\begingroup$ I see, now I understand your question. And one approach is to design a way for new labels that reflect class similarities. For example, one annotated instance with XL label is originally labeled with XL probability 1 and other labels with probability; the new label probabilities can be XL 0.8, L 0.1, and XXL 0.1, or M 0.01, L 0.09, XL 0.8, XXL 0.09, XXXL 0.01. A new model trained with new label probabilities can better reflect class similarities. $\endgroup$ – Tom Jul 25 '17 at 11:53

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.