The Music Explaura

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

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 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.


Post a Comment:
Comments are closed for this entry.

This is Stephen Green's blog. It's about the theory and practice of text search engines, with occasional forays into Machine Learning and statistical NLP. Steve is the PI of the Information Retrieval and Machine Learning project in Oracle Labs.


« July 2016