Apropos my previous entry on ZoomCloud, Paul Lamere posted a blog entry about a pretty cool little perl tool called Album Art Cloud (written by Andrew Hitchcock). This tool leverages Musicmobs’ open WS API to build a composite image of album cover art based on play counts of various songs. Here is mine (dynamic, browsable version available here):

Brooke's Album Cloud 2006.03.20

The cover art from albums listened to in greater frequency are displayed in a larger size, similar to how tag/word clouds display words with more frequent occurances in a larger font. In theory, this approach is a more interesting way to view the listening habits of a person or group of people. Instead of just looking at a bar chart or play count frequencies, one gets a visual map of album cover art, which one can mouse over to see play frequency, or click on to look at the album’s page on Musicmobs. What sites like Musicmobs, Last.FM, MusicStrands, etc. lack is good visualizations for their data. That’s why an Album Cloud is such a cool idea. It takes the a rich source of data and builds a great visual browsing interface.

As Andrew notes, there are some problems with actually implementing the cloud. Evidently, Musicmobs only has cover art (pulled from Amazon’s API) for a small percentage of albums tracked. If the Art Cloud program can’t find the album art for a particular album, it won’t visualize the frequencies for that album. This means that the cloud doesn’t accurately reflect play count. Compare, for example, my album art cloud above with a bar chart of my play counts on Last.FM:

Notice how some of my big favs of late are not present in the album cloud. Clap Your Hands Say Yeah, which I evidently listened to obsessively for a while, isn’t even in the Album Cloud. Ditto with Handsome Boy Modeling School. I am sort of comparing apples and oranges, since the Album Cloud is aggregated along albums, while the last.fm chart is at the artist level, but the point remains.

The problem here is that the cloud misrepresents a users’ listening habits — and more problematically, tends to skew the data towards more popular music. Popular music is more likely to be found on the Amazon WS API (depending on how the query is constructed), so less popular music, or music out on the long tail, gets overlooked.

There should be some way around having so much missing album art. Perhaps Musicmobs could straighten it out by refining their query against Amazon’s API. When we worked on Orpheus, we also used the Amazon API for retrieving cover art. Anecdotely, it seems like we were better at finding the art. But then again, we also used MusicBrainz to clean up users’ tags before submitting queries to Amazon. I’m not sure if Musicmobs does that.

A lack of album art on Amazon could become increasingly problematic as the music landscape changes. ‘Mash-ups’ generally don’t show up on albums, and aren’t sold on Amazon. Many smaller bands don’t sell their albums on Amazon, so the cover art isn’t there. As I said, if tools like this miss the long tail, they will become increasingly inaccurate. Perhaps an interim solution would be to just default to text when album art isn’t available. At least then the tool wouldn’t misrepresent the users’ listening habits.

Another problem with the Album Art Cloud is that is relies exclusively on play count to derive someone’s ‘favorite’ albums. There is other metadata that can help determine the users’ favorite music, such as ratings, repeat plays and appearance on multiple playlists. Since this data is available in the iTunes XML file, it should be available to Musicmobs as well, and thus to Andrew’s kickass tool. It would be nice to perhaps provide an option where users who rate their music (which I think is a relatively small pool of users) could include their ratings as part of the weighting scheme that determines the size of cover art. Thus, a song played only once, but rated a 5, could receive as large a rating as a song (or album) played 5 times with no rating. There are problems with this approach as well – Users have different ways of using the rating systems, and there is no consistent semantic interpretation for what the ratings actually mean. But it’s an idea. The other option would be to remove the semantic implication of ‘Favorite’ albums from the tool — instead calling it what it is — an Album Cloud based on frequency. This is what clouds typically represent anyway.

All in all, this is a cool little tool. Thanks Andrew!

On a related note, I just started using Musicmobs. It’s pretty cool. I especially like the playlist comparisons it provides, and the ‘similar listeners’ list. One can view other playlists or libraries that bare a resemblance to yours, see what songs you have, and which ones you don’t. In some circumstances you can even stream a section of a song that isn’t in your library.