3

Let's say I have a model which separates two classes with an SVM. Let's take a point from each of the classes. If two points have lower distance can this be interpreted as 'The model assumes these points are 'd' similar to each other ( low distance = high similarity )'. If yes, to what degree and how do I make sense of it in terms of my model accuracy?

Christo S. Christov
  • 133
  • 5

1 Answers1

3

Interpreting distance from hyperplane in SVM

Points that are further away from the hyperplane belong to their class to a greater degree. Identical distances express an identical degree of belonging (to their respective superclass) for two points, not a degree of similarity between two points.

MvZ
  • 146
  • 3
  • If you belong less to a class and it is a 2 class model then that means you belong more to the other class.. that was my main point of my question - is this true & how trustworthy it is based on accuracy – Christo S. Christov Oct 27 '17 at 12:46
  • 1
    Late, but worthwhile addition: It is always true if your classification construct is uni-dimensional (classify good/bad students using median test score as the only feature), but that hardly requires an SVM. What's to say that there is not a 3rd class that you do not know about, that explains some of the variance of features? I.E., an SVM will happily classify a 5-class set into 2 classes, but the feature weights will not represent the actual feature importance, and the distance is based in a nonsensical class definition. Suggested reading: Factor analysis, PCA, construct validity. – MvZ Sep 04 '18 at 10:31