Hunch homes in on who you are

It took 39 questions for the Hunch Twitter Predictor to make a wrong guess about me. The question was, "Have you ever ridden a Segway?" Yes, in fact, I have.

"We call it a 'stunt' internally. It's a fun way to show off the accuracy of our data," Hunch co-founder Chris Dixon told CNET about the Twitter Predictor, a new tool that takes a look at your Twitter network in an attempt to figure out as much as it can about you.

Start-up Hunch launched the prediction tool earlier this month and racked up about 20,000 visitors in its first weekend. Many of those who tried it out and, in turn, tweeted about it, had a similar reaction: It's scarily accurate.

"It doesn't analyze your tweets at all. We may do that at some point," Dixon said. "Imagine the Twitter graph as a little graph--so you're a point in the graph--and you have people you follow, and people who follow you. We have basically had a couple hundred thousand people come to Hunch and answer very detailed profiles about themselves. So imagine that the graph starts out blank, but hundreds of thousands of people come and fill out what we call their 'taste profile.' We let that taste profile flow through the graph, if you will."

So the Twitter Predictor, conceived in part by Ben Gleitzman, a recent Hunch hire from Google, takes your Twitter username. It looks at the people you follow, and the people who follow you. It looks at how many of those people have played around with Hunch's "taste profile" question-and-answer system. It connects the dots. And, for the most part, it works.

"We break people's taste into about 80 dimensions," Dixon said. "Let's imagine one dimension is political orientation, liberal or conservative, one is gender, one is food preferences, and each of those taste dimensions flows independently through the graph. Depending on who you're following and who's following you, we can make inferences about your food preferences or your political preferences."

The Twitter Predictor's accuracy is about 82 percent, according to Dixon. "It actually started off at 89 percent, but some people were claiming the questions were too easy, so in the middle of it we turned up the difficulty of the questions," he said. "We knew that would lower the success rate, but we thought that would make it more impressive." He also said that it's nearly impossible to nudge that accuracy much further than 90 percent in the first place. "People in our studies are actually only consistent with themselves about 90 percent of the time," Dixon explained, citing the accuracy ceiling that the developers trying to win the Netflix Prize competition eventually encountered. "The highest you could possibly give was around 92, because of the same thing. If you ask people to rate movies one day and the next day, they're only consistent around 90 percent of the time."

Plus, there are the inevitable quirks that it can't pick up: Hunch's Twitter Predictor blew it again when it asked me, "Do you know the signs of the zodiac in order?" and had every reason to assume that I wouldn't. It'd take an awfully fine-tuned algorithm to know that, even though it took me about 10 minutes and five mnemonic devices, I could remember the order of zodiac signs from way back in some comparative-religion class in college.

"It would expect you to know that answer because you were into astrology, and it probably thought you weren't into astrology," Dixon told me when I mentioned the Twitter Predictor's horoscope flub. (Insert joke about the relative accuracy of astrological predictions here.)

Finding legit uses
The New York-based Hunch, which Dixon co-founded with Flickr co-founder Caterina Fake, left some people scratching their heads at first. (CNET was the first to hear, way back in 2008, that Hunch was some kind of "collective crowd intelligence" recommendation engine.) Offering crowdsourced decision-making, the buzzworthy site soon started filling up with ways to find out the answer to "Should I quit my job?" or "How should I defend against a bear attack?" as well as a great way to procrastinate by filling out a "taste profile" with questions like "Would you ever set up a home theater on your own?" or "Have you ever been hit in the face with a pie?" It now pulls in about 1.2 million unique visitors per month.

Fun, yes. But the utility of the site wasn't immediately obvious, especially since I would assume that plenty of people are skeptical about turning to a crowdsourced Web site to learn how they should defend themselves against bear attacks. The Twitter Predictor is meant to show off how it can be legitimately useful--especially as Hunch prepares to launch an application program interface (API) that will bring its algorithm's know-how to third-party sites.

"We're releasing an API soon that lets anyone access our taste graph," Dixon explained to CNET. "So let's say you're using Foursquare or something. They know your Twitter name, and you want to get a tip if you just checked into somewhere in the West Village, you look it up...(or if) you go to Yelp, and instead of seeing reviews that are just random you can actually see reviews by people who have similar restaurant taste to you." E-commerce recommendations could be big, too.

There are two big challenges for Hunch, though: first, convincing more than just the early-adopter crowd who will try anything with big Silicon Valley names and venture backing (Hunch recently raised a $10 million round led by Khosla Ventures) that it's really worth their time to start answering stuff like "Were you a band geek in high school?" Dixon admitted that the Twitter Predictor is indeed more accurate for Twitter users whose network falls into that tech-crazed crowd. "It basically is better for people who are more similar to the typical Hunch users, which typically means early adopter types," he said. "Our weaker areas are in older demographics." Opening up the API and showing how Hunch can enhance services like Yelp or Amazon could make more people willing to start playing with it in the first place.

Second, there's privacy, something that could become an issue as people realize that, as with location-sharing services, they're turning over quite a bit of personal data. "We're very, very cautious about when we did this," Dixon said, adding that the company has been approached by marketers, especially in the wake of the Twitter Predictor's viral success, but has no plans to sell any data. "We're going to require that you can only get your 'taste profile' (on third-party services) if you connect to Twitter with OAuth. We're taking a very aggressive privacy stance even though, on a technical level, we don't need to."

Meanwhile, the company continues to tweak its decision-making engine, and has posted a blog entry about some of the odd revelations that the Twitter Predictor brought to light. Among them are the fact that some of the questions that it tends to get wrong are "Have you ever used a Polaroid camera?" and "Do you daydream?" The Twitter Predictor tends to inaccurately assume that quiz-takers don't daydream.

I don't know for sure, but I'd assume that people drawn to a service that lets them churn out 140-character bundles of thoughts are probably prone to daydreaming, too.