Playing with scores

Sorting is difficult, isn't it?

Playing with scores

Currently, there is a bug (https://github.com/distrochooser/distrochooser/issues/110) in the scoring method of the Distrochooser 5 beta.

The next Distrochooser "counts" in three dimensions:

  • Positive hits (P)
  • Negative hits (N)
  • Blocking hits ("no go") (B)
  • Neutral hits don't count at all

What is "expected"?

Lets first disuss how distributions should be sorted. An example:

  • Distribution A: 5 P 0 N 0 B
  • Distribution B: 5 P 1 N 0 B
  • Distribution C: 5 P 0 N 1 B

How should they be sorted? Of course, they may be in close reach as they all scored 5 positive hits. Distribution B scored in comparison to C no blocking (B) hit, but a negative (N) hit. As blocking hits are more severe, you would order them A > B > C. If B would have 1 blocking hit, it would have been A > C > B.

Playing around

Let's play around with some calculation methods on a sample result with 11 different distributions and a mix of positive, negative, but also blocking hits.

The scoring methods are basically variants from basic operations within the three scoring dimensions - just to see how they score!

In comparison, the different scoring methods show completely obscure results - as expected. Also, you can see how far the current (purple) scoring is from the "expected" method.

Position per Distribution and scoring method (lower is better)

Lets compare all scoring methods with the expected order. In average, the most proposals score 5 out of 11 distributions on the same position as the wanted one - with the only difference, that the scoring P-0,75*N-B scores 9 out of 11 distributions on the same position as the expected result. Thats good!

Lets compare the current scoring with that scoring.

The resulting order would have been D > A > B > C > E. In comparison: the current scoring would sort them A > B > C > D > E.

Lets compare this version with a slightly adapted version where negative hits have a slightly higher weight: The resulting order will be A > D > B > C > E.

This will cause the order D > A > B > C > E, same as with the previous one.

Conclusion

Take another look on this two scorings in comparison with the current and expected ones:

Position per Distro and scoring method (lower is better)

Both scoring methods are quite close to the expected one, even closer than the current scoring. Both with slight differences, due to the different weightings. I will play around with some weightings for N and B and will apply this within the next distrochooser updates.


Photo by Nick Hillier on Unsplash