Today I'm (finally!) announcing the first offering from
the AURA Project: The Music
Explaura. The Explaura is a way for you to explore musical
artists and find new ones that you might like, based on the words that
people have used to describe the artists. We call the set of words
used to describe an artist the textual aura for that artist.
You start out by searching for an artist that you know, say one of
your favorite bands. The data store contains information for about
30,000 artists. Over on the left, you can see what the Explaura knows
about one of my favorite bands, The Tragically Hip.
It's a bit hard to see (embiggened
version), but this gives you some idea of the information that the
Explaura collects for each band. There's a tag cloud (more on that in
a bit), the artist's bio from Wikipedia, videos from YouTube, photos from Flickr, album covers from Amazon and upcoming events from Upcoming. You can click on the
play icon to listen to that artist's radio at Last.fm.
On the left of the artist page, you see the list of similar
artists generated by the AURA recommenders. This list of artists is
generated using a technique that's quite a bit different than you're
probably used to. Rather than relying on the wisdom of the crowds via
a technique like collaborative
filtering, the AURA system computes the similarity between artists
by computing the similarity between their textual auras.
The tag cloud that the Explaura displays for an artist is a
portion of the textual aura that the system uses to compute the
similarity between two artists (in this case, it's social tags
collected from Last.fm.) This cloud is a
little different than the tag clouds that you typically see: here the
size of a tag is not proportional to its frequency, but rather to its
importance for this artist. Here's a better view of the
cloud for The Tragically Hip:
As you can see, The Hip are a Canadian band that plays energetic,
indie rock. How do we compute the importance of a particular tag in
the cloud? Using our good friend from the information retrieval
world, TFIDF. The idea is
that a tag is important for an artist if it is applied frequently to
that artist and infrequently to other artists (i.e., it does a good
job of distinguishing this artist from others.)
Because we're using the textual aura to compute the similarity,
it's easy to generate a set of words that explain the similarity
betwen two artists. If you click on the "Why?" link next to one of
the recommended artists, you'll be shown the overlap tag
cloud for these artists. Here's the overlap cloud for The
Tragically Hip and Sloan:
In this tag cloud, the size of a tag is related to how much that
particular tag contributed to the similarity between the artists. So
the fact that both The Hip and Sloan are Canadian played a pretty big
part in their similarity, along with the fact that they're both
literate indie rock outfits.
One more thing about the artist's tag cloud: if you click on one
of the tags in this cloud, you'll be taken to a page for that tag.
This page will look a lot like the artist page: it shows information
about the tag itself including the artists for whom the tag is
important. The tag cloud that is shown on the tag page is built from
the tags that are most similar to the tag that you clicked on. Here's
the tag page for classic rock:
But what if I want things that are like The Hip, but I don't just
want Canadian music? That's where steerability comes into
play. Each artist has a little steering wheel icon next to it. When
you click on that icon you're taken to the steering interface:
The steering interface starts out with a tag cloud that has the
most important tags for the artist. On the left, you see the artists
that the AURA system recommended based on their similarity to this
steering tag cloud. On the right, you can see a selection of tags
from the artist. Clicking on one will add it to the steering cloud.
Note that as you add tags, the recommended artists are updated in real
time. You're not restricted to the tags that have been applied to
that particular artist, either. You can search for tags to add using
the handy search box.
The really cool thing here is that the tag cloud is interactive: you can
drag a tag to increase or decrease its importance. If you drag
upwards on a tag, the tag gets larger and more important. If you drag
downwards on a tag, the tag gets smaller and less important. If you
drag a tag small enough, it goes negative and is shown with a
strike-through. When a tag is negative, no artists with that tag will
be recommended.
If we make the canadian tag smaller, then it's less important
and we get bands for which canadian is less important. We
can add the literate tag (because we like literate music!)
and make it bigger, which makes it more important. Again, the
recommendations are updated for each change in the cloud, so you get
direct feedback as to how your changes are affecting the recommended
artists. Here's my new steering page:
And there's a band that I've never seen before: Classic
Case. Now I can click on the play button and see if I like their
music.
If you don't want a tag in the steering cloud, you can right-click
on it and select "Delete" from the menu. If you click on "Sticky" in
that menu, then any recommended artists must have that tag in
their aura. You can click on "Negative" in this menu to quickly make
a tag negative.
It's probably a lot easier to see this in the demo video that Paul
made last year:
As you can see, there's been lots of updates since the video was
made, but there's still lots more to be done (for example, it's very
annoying that canada and canadian are considered to
be different tags), but we're pretty proud of how good the
recommendations are turning out to be. I've discovered several new
bands that I like using the Explaura.
There's a link for email feedback at the bottom of the Explaura
interface, so let us know what you think. I'll be posting more about
the Explaura and AURA in the future, so stay tuned.