Pandora Radio’s website says, “Pandora recognizes and responds to each individual’s tastes. The result is a much more personalized radio experience – stations that play music you’ll love – and nothing else.” But how do they do it? The project started in 1999 when Will Glaser and Tim Westergren founded The Music Genome Project, and within a few months joined Jon Kraft to found Savage Beast Technologies. The project was designed to break music down into categories and then establish music attributes for each category. There are over 450 listening attributes and five categories.
The Guts Of An Algorithm
In the months leading up to their 2011 IPO, Pandora talked about their methods and the algorithm used to program their online radio stations to respond to its user’s desires. There was a particularly revealing article in Fast Company Magazine that explained how the algorithm relied on far more than just electronic wizardry. There is an extensive team of musicologists working at Pandora, who’s sole job is to classify music tracks both new and old. These talented listeners use musical categories and listening attributes to breakdown each new track the company selects to list in their extensive library. Pandora pays royalties for the right to broadcast the music, and they want to only play music that a listener wants to hear. Thus the careful taxonomy.
Pandora Media is the largest Internet radio company in the world. It didn’t succeed just because it has a carefully worked out taxonomy. There’s much more to it than just that. If the service is not making great suggestions, listeners will only respect the recommendations a finite number of times, before moving on to something else. Also, the suggestions have to be made promptly because listeners are impatient and will only wait a few seconds for the next piece of music. Listeners need fast accurate information if the service is going to win their loyalty.
To accomplish speed and accuracy Pandora uses whatever information it can gather from a listerner’s previous choices. The music discovery and recommendation algorithm mathematically matches songs by crunching information it already knows about a listener’s choices and interests, matching them with music and artist attributes. But it’s the human ear of the musicologists that hear, identify and name those musical attributes. The idea is that the algorithm should make the listeners’ experience seem like Pandora is talking directly to him or her. As the suggested song is playing the listener can further refine the station by giving a “thumbs up” or a “thumbs down” to the particular song suggestion. The “thumbs up” teaches the algorithm more about the attributes the listener likes and the “thumbs down” teaches the algorithm about the attributes to avoid for this listener.
Other online services use collaborative filtering algorithms to make their predictions about what to offer. The collaboration uses the nearest-neighbor approach to compare. In other words, other listeners choices are statistically rated in relationship to the current listener and these comparisons guide the predictions of what else should be suggested. Pandora, however, is not interested in relating the interests of one user to another. Pandora only looks at comparisons of the current listener’s musical attributes drawn from their previous choices and then predicts by selecting from the Pandora music library’s song categories and attributes. The selection is calculated with nearest-neighbor techniques and a neighborhood of songs is built by arranging memory based data about previous choices into a neighborhood of musical category and musical attributes. Once this neighborhood is calculated it becomes the foundation of a calculated construction within the Pandora music library to create a similar, but expanded neighborhood of other songs. This is how the Pandora comparison is structured and the most similar songs are then rated as to how exactly they compare with the previously constructed memory neighborhood. The single song that matches best is selected as the next song suggestion for the listener. And it all happens in a split second, although the more “thumbs up” and “thumbs down” guidance the algorithm has to work with the more accurately it can project selections onto newly added musical content in the gigantic Pandora library.
The Business Deal
The company has selected both hardware and software that best supports the Pandora algorithm. The system is built largely on open source software and runs that software off a PostgreSQL database and uses the Hadoop distributed file system for offline storage. Pandora’s music files require so much bandwidth to deliver that most of the system structure is assembled around streamlining delivery. For this reason Pandora maintains its own private data centers,which are more cost effective than using public cloud infrastructure. Of course the algorithms are developed in house but the back end systems run on Debian distribution of the Linux operating system, which is also open source.
The year before Pandora’s Initial Public Offering (IOP) the company forged several important strategic partnerships. Music is listened to in cars. Half of most people’s listening hours occur behind the wheel, so it’s important for Pandora’s service to be part of the new car sales package in as many vehicles as possible. In 2010 Pandora struck deals with Ford, Mercedes, Buick, Toyota and Mini to include the Pandora online radio service. They also arranged deals with Alpine and Pioneer so that car radios can link with smartphone apps allowing Pandora to be played through a cellphone but over a car’s stereo system.
Pandora monetizes its service in two ways. First it sells advertisements that are played in between sets of songs. The advertisements aren’t played after every piece of music and they are only played to those who are using the free version of Pandora. For a $37 per year fee all advertisements are removed, although Pandora doesn’t push this second revenue method very hard because they are succeeding with the advertising revenue side.
In 2011 the Pandora 2.6 billion dollar IPO made six early investors very rich. Surprisingly, Tim Westergren, one of the founding partners, was not one of those pulling down a Midas sized pay day. The stock was initially offered at $16 a share, but quickly soared to as much as 63% in value before settling back to a 44% stock jump. Westergren owned about 2.3% of the company at that time and gained about $83 million dollars from the IPO. The two largest investors gained $800 million and $650 million respectively. Venture and angel capital pulled down the biggest IPO rewards.