@VannevarB      About      Pittsburgh Murals
June 16, 2009

Hunch Debut and Decision Making

Last month we saw the rollout of Wolfram Alpha, a computational knowledge engine (not a search engine, thank you very much). Today we see the rollout of Hunch.com, which is a decision-making engine.

The notion is that if Hunch can find out a few things about you, it can match you to a known profile, and help you make decisions by telling you what other similar people would choose.

So, you open a free account at Hunch. You answer a few questions about yourself. The questions are phrased categorically - for instance, they don't ask Do you live in Pittsburgh, in Allegheny County, or in Beaver County? They ask, City, Suburb, Rural?

Hunch asks you a few questions and develops a shallow profile of you. Every time you visit, Hunch asks you a few questions, generally in the this or that format that Optometrists use. For instance, the question: Do you prefer Fiji Water or Evian Water? tends to separate Republicans and Democrats. Every time you visit, Hunch's understanding of your profile gets a bit deeper, and theoretically the advice it gives you on decision making becomes more valuable.

Why are they providing this service for free? Are these altruists? From the FAQ: Some of Hunch's decision result pages display links to external sites, and some of those sites share with us a percentage of any purchases made by Hunch users. So if you ask for advice on insurance, you may end up seeing a Geico ad.

But the real point of the exercise is predictive analysis. This is how Amazon recommends a book and a music CD when you sign in. I've always wished that Amazon had a button for "I'm shopping for somebody else" so that when I'm shopping for Punk Nephew Billy-Bob, it doesn't skew the recommendations it makes for me.

The Hunch team has taken concepts from machine learning and applied them to human decision making, which is really a clever reversal of the usual process.

Predictive analytics needs a huge amount of information, and a large amount of interactions, to develop a neural net. Every time you interact with Hunch and get advice, you offer feedback - was this good advice or bad? Over a million operations, Hunch figures out how to give better advice next time. It's an idiot savant that needs us to teach it.

Once we teach it, it will be an extremely valuable property. That's why Hunch is offering you help - it needs you to teach the beast. Once the beast has enough transactions to get "smarter" and smarter, it'll be able to tell what you're statistically likely to decide, which is generally an economic decision - and there's a lot of money to be made in predicting economic decisions.

I've got a Hunch about this. (sorry)

Also: one of their key people is Pittsburgher (and Flikr co-founder) Caterina Fake. (wiki)




1 comments:

Anonymous said...

Not to be pedantic, but do they actually say they're using a neural net anywhere?

Neural nets were all the rage in CS about a decade ago, but have definitely fallen out of favor since then. The primary concern is that the damn things are totally inscrutable (in that their finished product looks totally arbitrary), and that the ability to learn an arbitrary function given exponential time is shared by other, less inscrutable, methods.

Post a Comment

Comments and Feedback? Love that stuff. Please leave your thoughts in the box below--