Applied Behavior Science for Health and Happiness
Pandora’s Siren Song
Pandora’s Siren Song

Pandora’s Siren Song

Pandora's Siren SongOne of the areas where sophisticated algorithms and deep databases have made particular inroads is in personalized music streaming services. There are a number of these on the market, but the one I know the most about is Pandora. While I’m not sure if what Pandora runs on is technically categorized as big data, it does seem to be able to learn from both deep and broad patterns to (fairly) successfully predict individual preferences, which is part of the promise of big data. And it does it with a focus on the what rather than the who.

Pandora’s approach to recommending music for people differs from other services that personalize offerings. Whereas a company like Amazon bases its recommendations on what people like you bought, Pandora focuses its analysis on the music itself. Through their Music Genome Project, they have isolated over 400 different variables that can describe a song. As one reviewer put it, “Pandora has no concept of genre.” Rather, recommendations are based on underlying features of music that may not be easily detectable to the casual listener, but seem to result in a high degree of user satisfaction.

I can say from personal experience two things: Pandora’s recommendations are pretty good, and when I listen to my custom Pandora station (cultivated over several years), I feel like Pandora knows me.  In this sense, Pandora has done a good job harnessing data to foster a sense of relatedness among users (something I think Amazon also achieves with its very different approach to data).

One thing that fascinates me about how Pandora uses its data is that it flips my implicit approach on its head. As a psychologist, I tend to think about the who first. Who is the user and what does this user like? Pandora starts with the what. What is the object being liked and what qualities does it have? There’s an elegance to the Pandora paradigm that I think might be worth exploring in other venues.

A problem that Pandora’s data approach could overcome is the redundant recommendation, something that is a huge issue for me with Amazon and with targeted advertisement in general. On Amazon, I frequently receive recommendations for items that yes, align with previous purchases, but which logically I am not about to buy:

Just bought a Thermos . . . do I really need another one?
Just bought a Thermos . . . do I really need another one?

And with advertisements, well, if I’ve purchased something, chances are you will have a low return on investment if you’re spending your advertising dollars showing me that same product over and over. Even The Onion knows that! I’d like to think that a more Pandora-like approach to advertisements would be smart enough to show me shoes that share features with the ones I’d bought, rather than latching on to the idea that I liked that one pair of shoes.