« Using the social graph to facilitate higher trust interactions | Main | Beyond traditional bank loans, specifically for small business »
Monday
Jan042010

On Second Thought, Hunch Revisited; Or, are preferences portable?

Addendum: Hugo has clarified our conversation, please read the comments.

Are preferences portable? Since my recent post about all the sites I love that track my tastes and make recommendations accordingly and my subsequent musing about the weakness of generic Q&A sites, I had the pleasure of a dinner conversation this weekend to shake all my thoughts up.

On the one hand, the former (sites like Pandora and Amazon) are niche and probably derive the power of their accuracy from that focus. The latter (sites like Hunch.com and Yahoo Answers) I found struggle to be all things to all people and fail in the process. I wondered: is it possible to combine the accuracy of niche sites with the breadth of generic sites? In other words, would bringing my finely honed Pandora music preferences over to Amazon help in their book recommendations which in turn could help with my Netflix movie recommendations? Over time with a successful universal ID that knows all my tastes, many decision engine questions could be rendered moot.

Hunch.com Chief Scientist Hugo Liu was not so optimistic. His experience as a "taste researcher" suggests that preferences are not portable. Amazon "knowing" my music tastes, he suggests, has a neglible impact on the power of their book recommendations to me compared with their data on what people who bought my most recently purchased book also bought. I wonder, then, why Hunch tries so hard to "get to know" you to inform their recommendations on a wide variety of topics. Hugo has discovered some pretty interesting, and perhaps revealing, correlations from the Hunch data, for example, that people who like to dance regularly overwhelmingly seem to also prefer Macs over PCs. Curious, yes, but how much should this discovery impact their recommendations?

More surprisingly, Hugo questioned the value of web personalization at all and suggested that it could even be dangerous. On New York Times, for instance, I am pretty happy using the "Most Popular" box as my guide to the top stories. Do I need personalized "Just for You" results? What would happen if Google started filtering my search results according to what I would be most likely to like? Personalization comes at what price of objectivity?

PrintView Printer Friendly Version

EmailEmail Article to Friend

Reader Comments (4)

Don't get me wrong, finding stuff I dig though recommendation bots es bueno, but sometimes I wonder it they're keeping me from discovery. I mean, I start to feel a little uni-dimensional when, no matter which Pandora station I'm on, Jack Johnson ends up chasing rodeo clowns (and then not too long after has the same chase acoustically). I'm starting to feel like Jack Johnson is to Pandora what chicken is to everything everyone's ever tasted... ever. How can I find something I never knew I liked? How will I discover my inner techo? Chances are it won't involve a rodeo... or Pandora.

January 5, 2010 | Unregistered CommenterKurt

melody, this is an interesting discussion for sure.
thanks for taking the time to post about it.

i would like to clarify some of what you took away from our discussion
as i don't agree with some of the points you report that i made.
my critique of the portability of preferences was directed specifically at
single-domain and single-application user action histories
for example, a vector of netflix ratings or amazon purchase history

in my experience, this kind of user model is too specific to the domain to be portable
they are quirky, fickle and possibly overfitted.
to use netflix ratings, for example, to make predictions about, say, a user's match.com matches
might get you 70% of the way there but you're going to introduce too much noise
a user action history is not suitable as a generalized model of preferences.

in my dissertation "modeling and simulating judgments of taste" (2006)
i argued for deeper, domain-independent taste-based models of users
not based on single-domain user action histories
but on a broader set of features about the person herself - her identity, personality, sense of humor, and attitudes

this is more or less Hunch's approach to preference modeling:
gather a broad set of features about the person
(try to avoid asking loaded questions, like "are you a good person?")
then, use statistical correlation to figure out which subset of these features
can be helpful for various decisions

for example, whether or not someone likes to dance might in fact be relevant to
whether they should switch from a pc to a mac
(aside: being a mac person myself, i am somewhat biased)
perhaps liking to dance is a good psychological proxy of people who like fun and embrace improvisation
(this is conjecture)

and creating a taste model makes a big difference to Hunch's results too
try training two accounts to different psychographics and see the Hunches you get without answering any questions
they will be very different for almost every decision!

finally, i do question the future of web personalization as you posted
personalization is hard to do right
and it is all too often a slogan and a gimmick in my field

you should personalize wisely, and recognize that it does have negative consequences
for example, google recently subjectivized their search so that different people will see different results.
given that rankings in google have become part of our consensus reality
("hey, just google for foobar and i'm the first result")
i am a bit concerned of the potential negative ramifications of this development.
but then again i may be proved wrong.

whatever the case, you can rest assured that Hunch will not sell you the fluff of personalization
i am personally too vigilant about making an honest product to allow this to happen!

peace.

January 5, 2010 | Unregistered Commenterhugo liu

@Kurt. You bring up some great points that I struggle with too. That's the danger of personalization that I see; that I only am exposed to ideas that I like rather than contrary ideas that may broaden my opinion. It's a similar problem to only reading bloggers that I like rather than a single newspaper that has a diversity of opinions.

January 5, 2010 | Registered CommenterMelody

Hugo, thanks so much for the thoughtful response and clarifications. Your research/thoughts on both preference portability and personalization have really made me re-think my thoughts in both areas. The distinction you raise is an important one -- maybe my Amazon book preferences will not easily indicate my Pandora music preferences, but a broader preference modeling (such personality through MyType) may help make recommendations across a range of domains. I'm excited to see how Hunch evolves, especially now knowing that you're behind much of it.

January 7, 2010 | Registered CommenterMelody
Member Account Required
You must have a member account on this website in order to post comments. Log in to your account to enable posting.