Joost Blog

« If You Just Can't Get Enough...Relive the Inauguration | Main | The Orignal "Lost" »

How Our New Recommendation Features Work

You may have noticed a few changes to our video pages last week. In addition to trying out a new design under the video player, we've just launched our (beta) recommendation engine - the "You Might Like" and "Have You Met?" bricks on our video page.

These "bricks" are "in beta" and will evolve over the next few months. Recommendation algorithms are tricky beasts, and you generally don't know how well one of the algorithms is going to work until you're using it with large production data sets and many users. We looked at a number of engines on the market and were quite impressed with what we saw, but at the end we decided to build this in-house so we could tailor it to the structure of our data.

I should really say recommendation engines, plural, as we're trying a hybrid of several different approaches that the team has been working on for the last few months. We've noticed patterns where people tend to travel in packs across our content library, and built our way from there to an approach that first recommends users with similar tastes ("Have You Met?") and, from that set, recommends content ("You Might Like"). We're not trying to be a dating service, but we do find that we often check out each other's profile pages when we're in the mood for something new. (For example, I tend to like Allan's music favorites and Samantha's comedy picks.)

When we don't have enough data to form a precise recommendation (which happens if you haven't logged in, for example), we switch over to an entirely different approach. For content recommendations, we're currently using an algorithm based only on what you're watching at the moment, a users-who-watched-x-also-watched-y approach. For "Have You Met," we blend in an editorial voice and select users from the community whom we feel represent interesting and diverse profiles at Joost. As a site with professional content, we are constantly playing with the balance of editorial versus data-driven approaches in our new features.

This is our first time out with these engines on Joost.com, and there we know there will be quite a bit of evolution. As always, if you have time to give feedback on these new features (or just the site in general), don't hesitate to write us with the good, the bad and the ugly at feedback@joost.com.

Posted by Kyle Forster on Jan 21, 09 |