Normalising the ratings to the rater average (or median, or mode)

[ilkka_nissila]

Bob, the point is not that individual people's ratings follow a normal distribution by default - they don't. If they did, there would be no point in the normalization. The point is that different people have different standards. The absolute ratings don't mean anything, because they follow different standards. Even though you may say that a 6 tells that someone has liked the image, another (a more trained person) might give it a 5 because they have seen much better images. Now, an image which is rated by a lot of people who know about art (instead of the novice) gets overrun by images which appeal to novices who rate only using extreme ratings. Why is the rating of the novice with strong feelings but no wisdom more important than that of the person who has had a lifelong training in art?

<p>

The only thing we know from an individual's ratings is that the person prefers the image which he or she gives a higher rating to an image that that person gives a lower rating to. <b>By normalizing, we extract this essential information from the person's ratings.</b> Which image does he prefer and how much over other images that they have rated? Once there are many ratings by different people, the end result is that the ratings average is 4, and 5 basically means that raters generally think it is a good image among those that they have rated. Which is precisely what we want to know. To get a 6 in the global ratings would mean that it's a truly exceptional image, because a lot of people have rated it higher than the average ratings they give.

<p>

I don't really care what the original distribution of a person's ratings is. It doesn't matter. Someone might want to excercise really loud voice by giving only the ratings 1 and 7. A more discriminating person might use 3,4, and 5 only. A third might look at the ratings given by others (which should be completely irrelevant to the rater) and try to pull down or push up the average rating by giving an extreme rating like a 2 or 7. These things happen. By normalizing the ratings by rater, we would extract the information from each rating and put all the raters on an equal footing. This would get rid of mate rating (because a person who only rates their buddies' images would be just walking without moving and he'd only be able to make a discrimination of which image he likes best of his friends' images - which is exactly the only bit of information which exists in the data given by the rater). It would also get rid of all sorts of extreme behaviour.

<p>

So you only want to give give ratings to images you really like. If you only give 6's - what objective information can the site draw from that? Nothing. 0 bits. A critique isn't a critique if all critiques are the same. -- I'll stop since I'm talking to deaf ears with a PhD.