Four Approaches to Music Recommendations: Pandora, Mufin, Lala, and eMusic
Thanks to MP3s and the Internet, we now have millions of songs readily available to us with the click of a button, but, paradoxically, this has often made it even harder to discover new music to listen to. Every online music store and every social network that focuses on online music, however, now features some kind of music recommendation system, and some services like Pandora or Slacker Radio are indeed nothing else but highly sophisticated music discovery engines. In this post, we will look at the different approaches behind some of the most popular music recommendation and discovery services.
Currently, we are seeing four different approaches to giving music recommendations in the market place - though the lines between them are often fluid and some services mash them up in different ways. For the sake of this post, we will only look at a small sample of music recommendation and discovery services that we think are representative of a specific approach.
Pandora: Humans Only
Pandora, one of the most popular music recommendation and discovery services on the Internet today, bases its recommendations on data from the Music Genome Project. The Music Genome Project assigns up to 400 attributes to every song. This, however, has to be done by trained musicians and the process can take up to half an hour per song. While the results of this method are often great, and we ourselves have often discovered interesting new music through Pandora, this approach simply doesn't scale very well and Pandora's library can often feel somewhat limited.
Mufin: Algorithms Only
Maybe the best known proponent of a music recommendations system that is purely based on algorithms is Mufin. Mufin's software analyzes the fundamental properties of a song and makes recommendations based solely on the musical similarity between songs.
While Mufin's approach generally works surprisingly well, the problem with this technique is that the system is simply oblivious to the cultural context of a song. Thanks to this approach, you might get to hear Christmas songs in February, for example, as the algorithms simply can't understand the cultural context of your music library.
At times, however, being agnostic to the cultural context of a song can also have its advantages, as Mufin's recommendations can often help you to rediscover music you had forgotten about. Mufin also works with any song, no matter whether it's from your own band, Kanye West, or an unsigned local band.
While Mufin's web service turned out to be a bit of a disappointment, we did like the company's iTunes plugin, which analyzes the songs in your library.
Lala: Explicit P2P Recommendations
Other services, like Lala, have decided to not feature any real recommendation technology at all. Instead, Lala purely relies on users following each other on the service and recommending new music to each other.
At least for Lala, this approach seems to work very well. When we talked to Lala's founder and CEO Bill Nguyen last week, he pointed out that 70% of all the music listened to on Lala was new music that was not already in a user's music locker, and that 18% of new music listened to on the service is bought and added to collections.
eMusic: Hybrid Approach
EMusic, the second largest online music store after iTunes, introduced a new recommendation system on its site late last year. This new system is based on technology from MediaUnbound, one of the larger providers of personalization and recommendation services. MediaUnbound, for example, provides the recommendations for MTV's Urge, Napster, and Brazil's Terra Sonora (eMusic, by the way, dropped Choicestream as its recommendation service in favor of MediaUnbound).
As MediaUnbound's CEO and co-founder Michael Papish explained to us last week, the company believes that a hybrid approach, which uses both algorithms and human input from experts, will provide the best results for users.
For eMusic, this means that the recommendations on the site are constantly fine-tuned by your own actions on the site, MediaUnbound's algorithms, and eMusic's editors, which, together with MediaUnbound's high-level teams, constantly evaluate the resulting sets of recommendations (Papish called this the "mosh pit" approach).
Genius: Apple's Black Box
It is hard to evaluate how Apple's Genius feature in iTunes really works, but Apple does have a few advantages. Because iTunes users often rate the songs in their library, Apple gets a lot of explicit information about a song's popularity. Users also regularly transmit information about how often they played and skipped a song to Apple's central servers.
Besides this, however, we can only speculate about what Apple looks at to give its recommendations. They surely evaluate playlists and the similarities between different users' libraries, for example. We can only assume that Apple uses a mashup of various recommendation techniques to come up with its own suggestions.
The results are generally quite good, though often either very predictable or completely random.
Opening the Black Box
In general, a black box approach similar to Apple's is still common for most recommendation services. Very few services give users a clear insight into why a certain song was recommended and the ability to fine-tune these selections (Pandora is a good example of a service that readily provides this kind of information). We are, however, seeing a trend towards users getting slightly more control over these recommendations. Slacker Radio, for example (see our review of their iPhone app here), lets users choose whether they want to hear more hits or more obscure artists on their radio stations. Mufin, too, gives users some control over how similar the recommended songs should be.
What Does Your Ideal Music Recommendation Engine Sound Like?
In general, we feel that every one of these approaches can provide us with relevant suggestions, depending on what kind of recommendation we are looking for. Of course, sometimes the easiest way to find great songs is to simply forget about the algorithms and editors and just look at what the people around you listen to on a service like imeem.
Do you have a favorite service for music recommendations? Or do you have some recommendations for these services that might help them to improve their service? Just let us know in the comments.
ReadWriteWeb Resources for Recommendation Technologies
We will be profiling other recommendation companies in upcoming posts. We also invite you to explore our custom ReadWriteWeb Resources:
- RWW Recommendation Industry Feed Favorites OPML file (save link)
- RWW Recommendation Industry Feeds - Best of Feed (copy and paste to your reader)
- Click to preview the above feeds before subscribing (pop-up window)
- RWW Recommendation Site Search (Visit and Bookmark)
CC-licensed logo image used courtesy of Flickr user shankar, shiv.
Share
Comments
Subscribe to comments for this post OR Subscribe to comments for all ReadWriteWeb posts
U should also include last.fm and blip.fm Both have a very different approach and seem to be working very well for me.
we feel that every one of these approaches can provide us with relevant suggestions..
the hybrid approach makes the most sense.
I think Pandora has a real opportunity to leverage all the great work that they've done with the Music Genome Project and make their algorithm the best in the business.
Computers are only as smart as we make them and Pandora has the ability to make the smartest music recommendation algorithm in the business.
Love all the services you've mentioned above Frederic but sometimes there may be a feeling of missing out stuff. To fulfill this "missing out" feeling, ReleaseDatez displays musics by their release dates and have the users create popularity based on how many people added the item to their watchlist. Take a look - http://www.releasedatez.com
The state of the art music recommendation comes from a combination of signal processing & pattern recognition, ie, recommendation based on content (digital files), rather than taggings or ratings. Content recommendation is closer to how human ears process audio signals, to sort out similar sounds/tunes/pitches/beats compared to non-content recommendation. There are some companies that I am aware that they're already doing this.
Content recommendation is still in its infancy, but it will get better and better over time and perhaps the content of the extracted features of songs will be added to the the Music Genome 400 attributes of songs as extra attributes. The more attributes in the dataset, the better the recommendation's accuracy.
I haven't seen this combination of song content extracted features with traditional song attributes (rating, tagging, etc,...) being done based on past & current researches, but I am sure that some researchers somewhere is working on it secretly at some secluded corner of the world. That would be a killer music recommendation app.
I use friend's radio stations on last.fm and ilike.com (facebook app).
great post.
last.fm is a big omission, no?
great post.
I hope that Pandora will be available soon in Europe:)
I really like Instinctiv Shuffle on the iPhone, but I have no idea how it works.
Interesting post about some of the different approaches; another new avenue of discovery however is 'social music recommendation' based on the psychological perception of music.
As you mention, "sometimes the easiest way to find great songs is to simply forget about the algorithms and editors and just look at what the people around you listen to." Taking that one step further, our new service Music Patterns provides customized playlists based on music that 'People Like You' actually listen to rather than simply from those 'around you.'
Using a psychology-based approach to music preferences , this method combines your individual preferences with identifying those that are similar to you and what they're listening to. Friends may or may not share similar tastes but using years of research in this area, best selling author Dr. Dan Levitin and our team at Signal Patterns offer a new way to discover music based on one's 'music personality.'
The state of the art music recommendation comes from a combination of signal processing & pattern recognition, ie, recommendation based on content (digital files), rather than taggings or ratings. Content recommendation is closer to how human ears process audio signals, to sort out similar sounds/tunes/pitches/beats compared to non-content recommendation. There are some companies that I am aware that they're already doing this.
And let's not forget Hype Machine = blogger-recommended music.
Pandora is great because the recommendations are song based (which helps since there are plenty of artists which I only enjoy x albums or x songs). Last.fm is great because it's recommendations (from what I understand) are essentially algorithms based on user data, so you get the human aspect. The strong points of these two services combined would be really powerful, but for now I have to sit on the fence between the two.
I have the same problem with all of these recommendation approaches: they are intended to find similar material to what I already like, confirming my prejudices and becoming a kind of "echo chamber" of style. While the ones that use more social algorithms (presumably the Apple Genius) reduce the problem, what I truly want is the new and different (and exceptional). Metacritic.com and a few choice critical web sites (some web-only, some old media properties) form the core of my new music discovery efforts today.
David Markowitz, your company's management team has got an impressive R&D background. There is the freely downloadable papers from some of the proceedings of Conference on Music Information Retrieval, just in case your R&D team haven't got those already (it is very likely they already have copies of those). Another useful resources is the Music Technology Group's list of available free papers from their site.
Something to ponder, if a company that develops a website for family genealogy gets $13 million VC funding lately (covered at TechCrunch/RWW last week) which doesn't involve brains in its development, then a company like yours must be worth VC funding of more than $50 million (first round only) because it involves real brains. It is a fact that developing complex stuff cost more than simple stuff.
I've been doing a lot of thinking about music recommendation engines on my blog. One concept that I've prototyped is a "piclens" type interface that let's the user hear samples of music just by hovering their mouse over a list of artists. This offers a rich user experience and also gives the algorithm a rich database by linking it to nuances of mouse movement. - you can check out the prototype here:
http://thinksketch.wordpress.com/2008/06/18/video-of-my-social-content-browser-prototype/
Let me know what you think -thinksketch
Just to clarify - in addition to the "P2P" recommendations in the Music Feed section, Lala does have system generated recommendations listed under the Albums You Might Like section of your homepage.
A really useful article, thanks! Personally, I see Pandora's Genome project as offering the best results, and this is something that users can refine over time, creating their own niches and a very personal playlist.
Whether this may also limit the ability to discover something unexpected and fresh, I don't know - is it possible to slacken the reigns on the results of your recommendations? I.e. 'gimme something energetic', be that techno or heavy rock. Maybe I'm assuming too much, and that the majority of listeners don't want to wander too far from their comfort zone. We'll see... it's exctiting times and, as I said, thanks for the useful post!
Lee Jarvis.
Thanks for the interesting post and comments. This is an important topic for our company, Emergent Music LLC, because we recently launched a music recommendation site, FlyFi.com, that is based on a client application we previously created and made available at Goombah.com.
Our goal at FlyFi.com is to make insightful recommendations for music lovers that can capture their range of taste in music, and enable them to discover music that extends beyond "mainstream" artists. For example, if you type in the artist Leonard Cohen into our application, our top recommendations are Tom Waits, Nick Cave & the Bad Seeds, Lou Reed, Nick Drake, and Jeff Buckley. Alternatively, Pandora's tup recommendations for Leonard Cohen are Bob Dylan, Tom Waits, Neil Young, Cowboy Junkies and Nick Drake, and Last.fm's top recommendations are Bob Dylan, Neil Young, Tim Buckley, Tom Waits, Joan Baez.
We feel that if you are familiar enough with Leonard Cohen's work to explore related artists, you probably already know about people like Dylan, Young, and Baez, so we go beyond these more mainstream artists to identify other artists that may be new to you. However, if you type a more mainstream artist into FlyFi, you can get more mainstream recommendations and recommendations that go beyond them, as well. Our goal is to enable people to explore music that incorporates their taste and is likely to be new to them, and that will perhaps even stretch their musical boundaries a bit.
Recommendations for artists the user types into our application are based on data mining and drawing correlations between artists. This leverages our proprietary approach for determining those correlations. We also provide users with the option to look at their entire music collection or portions of it, and/or a music profile they've created, as a basis for using our nearest neighbor algorithm to match them with users who have similar profiles and tastes in music. This provides a user with automated "word-of-mouth" recommendations from their matching users.
I almost always have Pandora running in a separate tab, but like you've pointed out, they're quite limited on their collection.
On the rare occasion that I'm not feeling up to listening to Pandora, I head on over to Last.FM, which turns out something that I may like every few songs, usually.
Then, on the extremely rare occasion that I'm not feeling up to either service, I head on over to iLike and listen to the fastest-rising songs radio or one of the other stations.
Don't rely on a recommendation engine, find it yourself with an accelerated immersive search.
Check out this prototype video of a cool 3D socialmedia browser-
http://thinksketch.wordpress.com/2009/02/17/cool-3d-socialmedia-browser-redux/
-ThinkSketch