2
$\begingroup$

I have a classification problem with three classes A, B & C.

I have features x1, x2 and x3 that represent my data items. But I also have a fourth feature that represents a similarity between my data item and each one of A, B & C.

My question is, how to encode this similarity feature given that it is relative to each of A, B & C.

My guess would be to have three similarity features sA, sB and sC. One for each category in the style of one-hot-encoding. So my resulting features are:

x1, x2, x3, sA, sB, sC

Questions:

  • Is this approach problematic in any way?
  • Is there an alternative?
  • Is there a name for this kind of encoding?
Improve this question
$\endgroup$
3
  • 1
    $\begingroup$ Is this similarity given already as part of the data, or you have computed it yourself starting from the initial features and the labels? $\endgroup$
    – desertnaut
    Oct 21, 2020 at 21:10
  • $\begingroup$ The similarity is computed from the features (but not x1,x2,x3) of the data item and the features of the label. The data item is a geometry and it has geometrical features. The label also represents a geometry which has geometrical features. I can calculate a geometric similarity between the data item and each label. I'd like to combine this with my 'regular' features (x1,x2,x3). $\endgroup$
    – totherwho
    Oct 22, 2020 at 7:53
  • $\begingroup$ And how are you gonna build these similarity features for the unseen data, where the labels are not available? $\endgroup$
    – desertnaut
    Oct 22, 2020 at 10:27

0

Reset to default

Your Answer

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