Musical Metadata
I’ve been using Pandora on my iPhone quite a bit lately. It’s great in the car when the radio isn’t delivering anything you are really into listening to and you’ve got a long-ish drive ahead of you. We also tend to spend long stretches one day during the weekend listening to Pandora via the home stereo while working on other things or reading. Given this amount of time spent with it; I’ve become somewhat intrigued by how it picks the songs that come up on any given playlist. First of all, I generally pick a playlist based on an artist as opposed to a single song, genre, or time period. Not surprisingly, there is overlap in my tastes and it makes a logical sense when listening to Leonard Cohen radio to hear Bob Dylan for example. Sometimes, we pick thing an artist just to see what playlist iwll come up. Some of these have been really wonderful and rewarding like Screaming Jay Hawkins but other artist playlists like Afrika Bambaataa somewhat disappointing. Lately, I’ve gotten into pushing Pandora to further limits of interrelationships trying to figure out how the metadata leads from one artist to another.
Lucky for me, Read Writer Web recently posted two videos featuring Tim Westergren explaining how Pandora works, how they’ve managed to gain financial feasibility and fore-shadowing the potential end-user charges that will be instituted in the near future. According to TW, each song in Pandora is assigned 400 attributes based on things like melody, harmony, rhythm, vocalization, etc. and these assignments are made by musicians listening and scoring each song. It takes 15-20 minutes for the standard pop song to be rated and over two hours for something like a symphony to be scored. The idea is to create a sort of “musical DNA” for each song. In addition to the 400 attributes, end-users are also assigning likes and dislikes via the thumbs up and thumbs down buttons on any given playlist. When you press a thumbs up, then that song gets assigned to that given playlist for the individual end-listener but also feeds into a larger overall approval rating of that song. There are some artists whose 400 attributes are so wide-ranging that it often times falls to approval or disapproval for their songs to end up in rotations. These artists include: Frank Zappa, Beck, and King Crimson.
My intrigue was initially developed out of the fact that I noticed when listening to The Smiths radio I kept hearing only two Psychedile Furs songs cycled in: “Love My Way” and “Pretty in Pink”. This seemed rather limiting so I went to the Psychedlic Furs radio and discovered that pretty much their major release catalog of musc was in Pandora but obviously not cycling high in other rotations. Now I know to favor these and see if they appear more distributed into other playlists. Last night, when stuck in a rather frustrating traffic situation about 20 minutes from my house, I typed in Tones on Tail radio and initially, did not even get a Tones on Tail song as a lead-off which is the standard playlist lead-in but instead started off with a Sisters of Mercy song. However, today, I’m getting better results so am thinking that maybe this was due to connectivity issues and not metadata relationships. I then tried The Chameleons UK radio and was pleasantly surprised to find that they had quite a bit of coverage of a band that never got much US radio airplay. However, there weren’t too many other artists appearing on this playlist. Within a 20 minute time frame, I heard four Chameleons songs and only had two other songs show up on the playlist: Joy Division and The Cure. This I think is due to the metadata but perhaps could be influenced by my likes/dislikes voting.
TW has stated that that the assignment of these 400 attributes makes songs “blind to popularity”. However, this is seemingly false statement if end-user likes and dislikes have influence in what shows up on any given playlist when the attributes cannot be solidly assigned as in the case of the three artists mentioned above. My interest here is because I’m drawing something of a parallel to something I’m seeing develop in libraryland. Library vendors are starting to incorporate some of these types of features into electronic management resources librarians use. In many cases, they are doing something similar, giving weight to certain attributes such as ISI rating, indexing of journals or articles, or authors along with retrieval rates from their systems, etc. These show up as a list presented to an end-user that says something like if you’re interested in this article you’ve searched, you may also be interested in these other twenty articles we’re presenting you with to consider. As with my personal interest in Pandora, librarians are interested to understand how the interrelation is being developed and the process used is not always spelled out to us or provided when asked about due to proprietary reasons. For instance, will these library-related tools weigh end-user feedback in a way that results in certain authors, publishers, and publications rising to prominence? Will there be, in part, a popularity contest occurring within the scholarly landscape or will new top journals emerge? Is there a way to use or analyze metadata to try to determine causation of patron scholarly e-journal usage and help us move beyond our current reliance on correlative data?
2 years ago