danielknox.net

During the Google TV announcement yesterday, Google mentioned that will be partnering with an Israeli startup called Jinni to provide recommendations based on an intelligent ‘taste engine’ for movies and TV shows.

More on Techcrunch here.

My colleague, Neil Budde just wrote an interesting article on the Newstogram blog about Facebook’s Open Graph and their Network Activity or Recommendations modules.

Neil identifies one of the challenges of using the Facebook Open Graph for making high-quality (i.e. relevant) content recommendations:

The biggest drawback to Facebook seems to be that it’s dependent on users clicking the “Like” or “Recommend” button on stories. We’ve been looking at Facebook widgets on a few of the large news sites, such as ABCNews.com, Time.com, CNN.com, and WashingtonPost.com. Even with up to 700 Facebook “friends” spanning all ages among our different accounts, we have few (or none) who are active recommenders of news stories. On a few of the sites, the only friends recommending stories are employees of the site, and even they aren’t that active. The result is a module not unlike a Digg or Most Popular list.

Interesting that the top two online retailers in the Foresee Online Retail Satisfaction Index (Netflix & Amazon) are the ‘poster children’ for e-commerce personalization. Would be worth investigating how many of the top 100 have implemented some form of personalized recommendations on their sites to see if this is indicative of a larger trend.

I gained some insight today into why (at least for me) Google Reader’s “selected just for you” recommendations suck (see previous rant here).

In a paper presented at the IUI ‘10 Conference a group of Google researchers discuss different recommendation algorithms they have tested on Google News. They outline some of the problems with their default ‘personalization’ method (which relies on collaborative filtering), including the inability to recommend new stories and the inability to account for variability between users leading to “recommendation convergence” (this is my term, not the researchers’, but I think its appropriate). For instance, they observed:

… that entertainment news stories are constantly recommended to most of the users, even for those users who never clicked on entertainment stories. The reason is the entertainment news stories are generally very popular, thus there are always enough clicks on entertainment stories from a user’s “neighbors” to make the recommendation.

I assume that Google Reader is also using a collaborative filtering method to recommend articles “just for me” since the recommendation convergence issue would definitely explain why all I seem to get recommended are humorous (and presumably popular) videos.

I can only hope that the hybrid approach that was tested on Google News (and which performed 30%+ better than collaborative filtering) will be rolled out to Google Reader as well. Until then I’ll have to put up with clips “selected just for me” (and thousands / millions of other people ‘just like me’) like a lightning blot striking a plane, a baseball player jumping over the catcher and a young girl doing a trick on a bicycle (actually that last one is pretty cool!).

The Filter, which recently announced a partnership with NBC.com to power video recommendations, got a write-up in this week’s BusinessWeek.

Included was this excerpt from The Filter invester Peter Gabriel which shows The Filter’s messaging and ambition is very similar to Hunch.

Gabriel says he’s certain The Filter could one day serve as an all-purpose “decision guru” that goes beyond music. He sees it as a tool for people who find themselves in, say, Barcelona and need quick tips on where to dine and how to dress. “I think the same way we got used to Google being part of our lives and asking it questions, The Filter will be like our own Babel fish [the translating creature from the comic novel A Hitchhiker’s Guide to the Galaxy] who will help us make decisions,” Gabriel says.