I know this topic is dead, but just adding a note about the distribution of ratings in a library in response to the previous comments.
Ratings should absolutely NOT follow a gaussian distrbution unless the library is very small, but large enough that it contains a random assortment of music such that there is a small amount that you haven't listened to much, a small amount that you have listened to a lot and most have been listed too somewhere in between.
Play counts will undoubtably follow a Poisson distribution.
http://en.wikipedia.org/wiki/Poisson_distribution...where the expected value (Lamda) is the average play count (or frequency) of the library and "k" is the actual play count of the track. If you listen to your music a lot and the average play count grows faster than the average length of time in the library, the curve will begin to look more like a normal distribution, where as when the opposite tendency occurs, it will tend to look more like an exponential decay curve.
If we assume a linear relationship between how often you play a song and how much you like it, which I think is fair*, then the ratings should follow suit.
Chances are this is why the distribution of ratings many people get will be heavily skewed to the lower ratings. For the simple fact that if you have a large library, and a song isn't one of your most favourites, then the probability of it being played any given time you listen to music is low, keeping in mind that there is a random interval of time between subsequent plays of songs, which is relatively true universally.
I've spent some time attempting to utilize this as a possible alternative method of determining a rating based on the probability of a track having the rating it does, but that's not going to be in version 1.5, but maybe subsequent versions.
To [somewhat] correct for this phenomenon, I've added the "second order optimism factor" to the algorithm which essentially skews results towards middle values in an upward direction only.
