Latitude

A recent article in New York Times Magazine offered some fascinating insight into the theory and process behind Pandora’s human-supported recommendation engine, with implications for purely “experience”-focused recommendations vs. ones that rely heavily on tiers-of-interaction data & collaborative filtering.

How Pandora Works

Pandora’s service utilizes variables outlined in the Music Genome Project to score nuances in instrumentation, vocal intensity, musical genre, tone and content of lyrics, etc. for every song in its database–which then can be used to create a personalized, self-refining algorithm for each user, recommending novel content based on “predictable” preferences.

(In the Music Genome Project, there are nearly 400 possible variables; about 150 of these are relevant to rock and pop songs–more for rap and jazz tracks–with each rock or pop song taking about 20 minutes for a highly trained human to code.)

On the research front, we think that Pandora is an interesting mainstream example of a human coding project–that is, of projecting structure (a “code” composed of music-related variables in this case) onto otherwise “qualitative” or unstructured content (a song) for the purposes of analyzing this content along self-derived axes, elucidating meaningful relationships between variables–and, often, creating an algorithm to predict future outcomes.

Creatively conceived, “coding” can be superimposed on most any information inherent in observable human behavior, music, content–or even the way we perceive a visual image (ex. think metadata).

Pandora Doesn’t Share

In a subtle uprising against Web 2.0’s oversharing tendencies (insofar as they affect and obfuscate our own natural preferences), Pandora actually tries to minimize the role of social in its recommendations to tap into our personal, unadulterated experiences of music.

Pandora’s approach more or less ignores the crowd. It is indifferent to the possibility that any given piece of music in its system might become a hit. The idea is to figure out what you like, not what a market might like.

New York Times Magazine, “The Song Decoders at Pandora”

Pandora vs. Genius’ Purchase-Driven Recommendations

The best we can figure that one recommendation competitor, iTunes Genius, does its magic (both within your own library, and in calling up track suggestions for purchase) is through direct engagement signifiers, prioritizing the kinds of music you and other users like you (based on your library contents) have purchased in an attempt to–well… drive more purchases.

Westergren [of Pandora] maintains “a personal aversion” to collaborative filtering or anything like it. “It’s still a popularity contest,” he complains, meaning that for any song to get recommended on a socially driven site, it has to be somewhat known already, by your friends or by other consumers.

New York Times Magazine, “The Song Decoders at Pandora”

Experience “Data” vs. Purchasing Behavior

“There’s no rating that allows an analyst to conclude that a vocal solo is simply lousy…”

A purchase represents, perhaps, the most involved end of the engagement spectrum. But, from the experience standpoint, what I purchase isn’t always what I truly like best. For example, I may not have purchased any of the artists that my roommate currently has in his library because we can share them over our network. (And from the purchasing behavior standpoint, that doesn’t mean I won’t buy these artists’ future releases.)

“What Pandora’s system largely ignores is, in a word, taste.”

Pandora, blocking out social data and, instead, working from “the bottom up”–tying users’ “yea!” or “nay!” reactions to intrinsic elements in the music–serves as something of a control group for experience-based recommendations.

“There were elements of music that machine listening just couldn’t capture — like the emotionality of a Getz solo.”

If I answer “yea!” or “nay!” ad infinitum, ideally, Pandora will learn which variables seem most important to me and how I tend on their scales (though these are obviously subject to context and situational factors).

It would be nicer if I didn’t have to begin tabula rasa–if I could select “priority variables” alongside my artist or track “seed,” and have Pandora work intelligently as it does from there.

But iTunes Genius never asks for real-time engagement in order to self-refine (it doesn’t “get smarter”), though supposedly it considers any song ratings I’ve previously attached when serving up recommendations. I think it’s reasonable to assume that iTunes recommendations (for purchase) are optimized to incent “conversions.”

Does it make a difference? Endlessly attentive and malleable to your nuanced reactions, does Pandora actually provide a better experience?

Header image courtesy of marfis75’s flickr, (cc) some rights reserved.