is an interesting experiment in movie recommendation.
“Our approach is new and different. We write magical search algorithms that interpret comments people like you have written about movies. No editors. No movie critics. We analyze the millions of viewer comments from all over the Web to gain exciting insights into what movies are really about and what people think of them. Based on those insights, we find movies that you’d like to watch.”
Doesn’t Netflix already have this type of function in use and has it based on the ratings you give movies, TV, and even genres entirely?
They’re a little different in that they draw their information from different pools (or silos, choose your metaphor).
Netflix pulls only from Netflix users, then it compares the movies you’ve rated with their broader rankings. The average star rating for the latest action-adventure flick may be three stars, but Netflix will predict that I’ll give it 4 stars since I regularly rank action-adventure flicks highly.
Nanocrowd scrapes reviews from across the web and then generates five sets of keywords. So in the scraping it collects a set based on genre (action-adventure). Once it has that set it asks me what I liked about the movie and gives me five options – explosions, fast cars, guns, suspense, exposed flesh. When I select one of those five options it generates a subset that combines fast cars with action-adventure. It generates “fast cars” because a certain number of Internet reviews referred to fast cars within the review.
Netflix is actually offering a $1 million to someone who can improve their recommendation algorithm. I’m not sure why they don’t simply expand to a 7 star system which would allow for a lot more granularity. What they’ve done instead is create a ton of new categories and ask for you to rate each category.
On the one hand I love recommendation systems like this. Amazon recommendations have led me to some great books. OTOH, one of the things I love is novelty and the unexpected. Recommendation systems tend to create sets that are too narrow for my liking. They need to include an orthogonal set as well, along with “you’ll probably love,” they need a “you’ll probably hate,” or “you’ve probably never even thought about” choice.