Re: Sorting scores



Bruno Barberi Gnecco <brunobgDELETETHIS@xxxxxxxxxxxxxxxxxxxxx> writes:

I've been looking for ways to sort scores. It's not a rare
problem: in a scale of 0 to 10, if one uses simple average an item
that has only 1 score of 10 is ranked below an item with 5 scores
of 10 and 1 score of 5. I suppose it should be also meaningful for
a score weight to decay with time.

I'm sure this problem must have been studied before, but I
can't find any references. Thanks a lot for any help.

There is no unique best way to do this. One way is to assign a
confidence to every rating, which starts out at, say, 10 and decays
over time. You can also rate your reviewers so those you trust the
most start at a higher confidence and decay more slowly.

So the question is: If you have a set of ratings/confidence pairs, how
do you combine them into a single final score used for ranking?

One possible solution is:

1. Let C be the sum of all confidences.

2. Let R be the sum of rating*confidence for all pairs.

3. Let score = R/C.

This way, if all agree on the same rating, the score will be equal to
this rating, regardless of their confidence. It is still not quite
what you want, as 5 ratings of 10 plus one of 5 scores lower than one
rating of 10 (regardless of their confidences). So you would want to
give a bonus to having more ratings and higher confidences.

One way to achieve this is to add a single artificial review of
average rating and high confidence to the mix. This way, it will
require more reviews to pull the score away from the average.

Let us, for example, assume that 5 is an average rating and
confidences go from 1 to 10.

Example 1 has a single review of rating 10 and confidence 10.

Example 2 has five reviews of rating 10 with confidences 10, 9, 8, 7
and 6 plus one rating of 5 with confidence 10.

When we add the artificial review (5/10) to the first example, we get

C = 10+10 = 20
R = 10*10 + 5*10 = 150
score = 150/20 = 7.5

When we do the same to the second, we get

C = 10+9+8+7+6+10 = 50
R = 10*(10+9+8+7+6) + 5*10 = 450
score = 450/50 = 9

I hope this is at least a little bit helpful.

Torben

.



Relevant Pages

  • Re: FEX submission reviewers under close scrutiny
    ... simply fail to learn from constructive criticism. ... it is truly the right of reviewers to give any rating they wish, ... simple peer review system. ... a "poor" rating, ...
    (comp.soft-sys.matlab)
  • My reads for April
    ... My rating: Too long on action, ... For full review: http://sunniesbookblog.blogspot.com/2009/04/review-greasing-pinata-tim-maleeny.html ... Its about time there was crime fiction written from the perspective of a coroner who serves as both law officer and investigator. ... Detective Emmanuel Cooper is sent to look into the murder of an Afrikans Police Captain in a country town. ...
    (rec.arts.mystery)
  • WWE Friday Night Smackdown International. Feb 29 2008 Review
    ... My Review of this weeks Smackdown. ... WWE fans exactly what kind of condition the WWE is actually in at the moment ... Rating? ...
    (rec.sport.pro-wrestling)
  • Re: Sky & Telescope is now starless?
    ... Suppose you're rating optics. ... best Newtonian will end up with 3 or 4. ... Newt and an APO of the same size? ... I also read a review that offered faint praise for ...
    (sci.astro.amateur)
  • Re: Why is there such a disconnect?
    ... John Harkness wrote: ... many critics abhor having to come up with any ... hard to imagine a detailed review that might give a reader only the ... your erstwhile buddy Mike D'Angelo uses a 100-pointrating ...
    (rec.arts.movies.current-films)