0
$\begingroup$

I have some data similar to movie ratings and the labels are ordered, like 1 to 10. since the target label is not a nominal but ordinal variable, what types of models should I be using for classifying such data? what kind of losses and metrics to use?

$\endgroup$
3
  • 1
    $\begingroup$ why do you think the loss functions or the models should be different than other classification tasks? $\endgroup$
    – Nikos M.
    Jun 11 at 16:48
  • $\begingroup$ because the data is not usual, it is ordinal. and there is relation between each number. i learned that weighted-kappa can be used as metric, but don't know about loss and algos. $\endgroup$ Jun 12 at 4:25
  • 1
    $\begingroup$ The fact that the target is ordinal does not make the models or the loss functions different in general $\endgroup$
    – Nikos M.
    Jun 12 at 9:11
0
$\begingroup$

In classification problems one usually uses categorical variables. An example are One-hot vector, that have a 1 in the index of the corresponding label and 0 on the rest:

label 3 -> [0,0,1,0,0,0,0,0,0,0]

So if you transform your label to a one hot vector, you can now create a mathematical model.

This is accompanied by a softmax layer at the end of your model to represent the probability that a given input belongs to each class. You can compare this output vector with the one hot vector from the ground truth using, for example, cross entropy loss. This last link shows a nice post that explains the whole procedure.

$\endgroup$
1
  • $\begingroup$ one hot encoding is killer regarding dimensionality increase. Instead in many cases other encodings can provide state-of-the-art results. However in this case the target is already encoded $\endgroup$
    – Nikos M.
    Jun 11 at 17:52

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.