Friday, August 10, 2007

What we know about Pandora

How exactly Pandora works is, of course, proprietary information that helps keep Pandora in business by selling ads and subscriptions. Nevertheless, Tim Westergren (CEO) and Tom Conrad (CTO) have revealed bits and pieces of how the player works some of which is not widely known.

The Genomes The simple way to understand the Music Genome Project is that there is a set of traits which are assessed for each song in the database. More accurately, however, there are several such sets. The fact is that the initial genome did not make useful distinctions for certain genres like Electronica and Rap, and so they built more detailed genomes tailored to these genres as they entered into these genres. There is no genome for Classical music, and it remains unclear how high on the list of company's priorities is the creation of a Classical genome. Or, perhaps, it's proven particularly hard to create an effective genome for Classical.

The Focus Traits The traits listed on most song pages are NOT the genome traits. The listed traits are called "focus traits" and are calculated for each song after the genome traits are assessed for a track. The evidence is fairly clear that there is a different set of focus traits for each genome. For instance, Electronica tracks tend to list many more focus traits for each song than, say, a rock track. It also should be noted that around 5% of the songs which play on Pandora list no focus traits at all.

Song Assessment Pandora buys CD's to acquire the license to play the tracks of that CD on the player. However, when one of the trained Pandora musicologists tackles a CD they do not, in general, assess the genome for every song on the CD. Thus, even though a track may have a page in the database and even a sample, the track may not be available to play on Pandora because it's genome has not been assessed. If and only if a song lists "Similar Songs" on its page will that track be available for play on Pandora. The company will go back and add a song if enough people search for it; however, if one of your favorite tracks was passed over when they first covered a CD and that group does not have a following on Pandora, then it's unlikely that the track will ever appear on Pandora.

Song Selection The non-Quickmix player generates songs in sets of 3 or 4. The player appears to select a "keystone" song for a set from the list of the seeds (the Artists or Songs you selected for the station), the list of thumbed-up songs, or (rarely) one of the songs which are considered "similar" to one of the songs in those lists. The keystone can play in any position in the set (1st, 2nd, 3rd or 4th), and the other songs in the set will be particularly similar to that song (often these songs will appear in the list of "Similar Songs" for that song on the keystone's song page). Since June 20, 2007, the song selection algorithm has increased the chance that songs from the lists of seeds and thumbs-up will be among the additional songs in a set, and so it is more difficult to identify the keystones when listening. It appears that when a Quickmix plays several songs in a row from the same station, the songs are from an identically generated set of 3 or 4 songs.

No comments: