experiment

FastCo Labs article reveals how Pandora continuously experiments and tweaks programming

Friday, August 16, 2013 - 12:15pm

Fast Company Labs has a fascinating article that reveals the extent to which Pandora experiments and researches how its audience reacts to different variables in the way it creates its playlists -- with the aim of increasing the rate at which users return to the service.

[By the way, if this sort of thing interests you, please our coverage of Rhapsody VP of product-content Jon Maples' on the importance of music curation here.]

We really want to encourage you to read the entire piece, but we've pulled out some bits we found particularly fascinating.

According to the article, by John Paul Titlow, Pandora's data scientists regularly divide and subdivide its audience into test groups, then continually tweak how music is delivered to listeners. For instance, they might vary how often songs are repeated, or the ratio of very familiar tunes to new music. Perhaps they'll vary the concentration of artists that are "local" to a listener, or how many "live" or "acoustic" versions of songs a listeners hears. They even monitor how listeners react to music given their geographical location, or the time of day.

Pandora has run thousands of these tests over the years, some months-long, some taking just a few weeks. And they've apparently resulted in some very interesting insights. For example, the webcaster has found that listeners are less tolerant of unfamiliar music while they're at work. So the webcaster has adjusted for this, and now your personal Pandora channel may seem more familiar between 9 and 5, and a little edgier at night or on the weekend.

Or, fans of instrumental music (like most Classical and Jazz) are generally more receptive to new music discovery -- fans of vocal pop music, the opposite. Titlow writes, "The distinction is so pronounced that stations based on instrumental hip-hop will yield more serendipitous moments of discovery than those based on lyric-heavy rap tracks."

Pandora has even tracked how the same listeners may interact with the music differently based on which type of device they're using at the time -- on the web, or a mobile phone, or a Blu-ray player in a home theater.

While much has been made about the origins of Pandora's Music Genome Project -- hundreds of trained music expers dissecting each track and scoring it on dozens of characterists -- it's user data (skips, "thumbs up/down," etc.) that are training the system now. In fact, Pandora listeners create data far faster than its staff of human experts can. And to be able to more quickly ingest new music, Pandora has developed its own "machine listening technology." It merges the computer analysis of music with input from human experts "to create a deeper understanding of the music its service spins."

The article ends with a short bit about applying this intelligence to the dynamics of group listening, and how new technology could enable that. Again, we'd like to encourage you to check out the article here.

Syndicate content