I have to say that while the aim of the article is quite admirable, the author does a horrible job of analyzing the data he collects, and he completely misses an anomaly of interest.

His conclusion that "only for tracks rated 1-19" do ratings of finer granularity matter is bogus. Ignore figure 5 and look at every other figure in the article. It's not linear from 1-19 either. It shows the same step behavior as the rest of the graph. He mistook an anomalous edge case for some sort of liner relationship. The *actual* interesting piece of information that can be taken from the data collected is that, whether intentionally or not, the algorithm Apple uses has an odd (again, I don't say erroneous since it's possible that it's intentional) edge case. The play count difference between two songs where one has a rating one star higher is around 2000-2500. However, between no stars and one star, the rating difference is 3700. Without more information (and a Mac of my own to investigate myself), it's not clear what's going on. The author states that the 0-100 rating gets integer-divided by 20, which makes sense except that that's not what the data in the graph reflects since a rating of 0 yields a significantly different playcount from a rating of 1. In fact, neither a strict ceiling or floor explains the data generated, especially since the rating is discretized to a five-star rating.

Here's my point in summary. If iTunes integrally divides or floors the rating, then where does the extra step when the rating is between 1 and 19 incluseive come from?