How Hunch Works

For a normal person's overview of what Hunch is, see our about page. But if, like many of us, you're the type of person who likes to watch those "how it works" shows, here's a little more detail that might appeal to your inner geek.

The theory behind using collective knowledge for decision making

Researchers have documented how decisions made by diverse and independent groups of people are often superior to those made by individuals - even experts. The reason is that knowledge is often spread among many people. The challenge is to identify it, collect it, and effectively use it.

Take, for example, expertise about colleges or cars. In a random, large group of people, most probably know something about a few examples (say, the college someone attended or the car they currently drive) but are not experts on the topic as a whole (as a college guidance counselor or auto executive might be). If you were able to collect and organize all the various bits of individual knowledge that the large group possesses, you’d have a pretty complete picture of the topic overall.

The theory in practice: The Hunch algorithm

At the core of Hunch is a question selection algorithm built by our small gaggle of MIT computer scientists with backgrounds in machine learning. The algorithm is always asking itself, "What can I ask you next which will lead to the best possible recommendation for this topic?" The choice of which questions to ask and when to ask them will vary based on what you've already been asked (and how you've answered) so far, the same way that a human expert would adjust a line of questioning based on your responses. The idea is that if someone says they're a vegetarian, you don't want to then immediately ask them how they want their steak cooked.

In choosing what to ask you, Hunch's question selection algorithm tries to do two things. First, it tries to find a question which will discriminate well among the remaining possible recommendation outcomes for you - thus filtering the remaining choices from "many" to "fewer". Second, the algorithm looks for a question which can help optimize and rank the remaining recommendation outcomes to present you with the ones you'll like the most. It's trying to ensure that you'll like outcome #1 better than outcome #5.

User contributions are mostly what makes Hunch smart

As you answer questions, Hunch can narrow down your possible recommendations outcomes because each outcome can be "trained" to correspond with each question's answers. Any logged in user can set initial training or correct existing training, in addition to proposing new topics, questions to ask, and recommendation outcomes. This is how Hunch is truly a collection of common knowledge. So whether you happen to know a great question that would lead someone to a Sancerre vs. a Pinot Grigio, or you'd like to clarify that "Whatever Happened to Baby Jane" is probably more of a "campy" than a "cult" movie, Hunch absorbs your input and uses it to provide smarter recommendations for the next user.

Hunch also uses machine learning based on statistical inferences

Besides users explicitly training and contributing to Hunch, there's a second way that Hunch learns, especially for what we call 'Teach Hunch About You' questions which have more to do with you as a person than with your preferences for a specific topic's objective filtering criteria. When a user clicks "Yes" or "No" to indicate whether or not they like one of Hunch's recommendations, Hunch incrementally strengthens or weakens the mathematical correlation between that recommendation and any 'Teach Hunch About You' questions that have been answered so far. So over time, Hunch might learn that people living in cities tend to prefer diet sodas, or that SCUBA divers tend to like bicycles with lots of gears. (we just made up those examples, but you get the idea.) The academic name for this sort of algorithm is machine learning.

So, like we said...

Hunch is designed to soak up collective knowledge and then organize it in a useful way to offer you smart recommendations. Hunch proposes custom recommendations for you that it wouldn't necessarily give to somebody else. But at its core, Hunch's recommendation algorithm is just a mathematical framework. It's the users of Hunch who give the algorithm proper training and personality by contributing to it and making it clever, funny, and nuanced.... but most of all very useful in helping everyone to get smart, efficient recommendations.

For more information about the specific features of Hunch, check out our FAQs. If you're still befuddled, take a browse through the Hunch forum.