my personalized shopping experience is the same as yours
March 7th, 2007
Amazon has recommendations, LibraryThing has the LibrarySuggester and other sites have other variations of this, but they are never quite right. Why is that? Do we not find the recommendations credible?
Amazon makes recommendations based on the book you are viewing, so if you like Harry Potter, you might also like what others who bought Harry Potter purchased. However, just because I purchased something doesn’t necessarily mean I liked it. LibraryThing makes recommendations based on a single book, which has the same limitation as Amazon. For members, LibraryThing also makes recommendations on all the books in your profile, matching them up against others with similar libraries - I assume your book ratings are taken into consideration. Surely this is as good as it gets?
Or do we need to go one step further and ‘decode’ a book the way that Pandora decodes music? Unfortunately, we all know how hit or miss Pandora recommendations can be. Along with Last.fm, Pandora’s recommendations are probably the best I’ve seen, but they still aren’t perfect. Maybe it’s easier to decode a book than it is to decode music?
I feel as though we can do better.
March 15th, 2007 at 5:38 pm
We can do better. Netflix is the master here - they gather more data, and use it better than anyone, even Amazon. Amazon generally knows who bought what book, but Netflix asks EVERYONE who rents a movie through them how they liked it and apply that back to their algo. They know this, and that’s why they’re paying $1M to someone to make their algo better.
Unfortunately us in Soviet Cannuckistan can’t use their great service. Come on Reed, we can help you grow by 10%!
That being said, the problem is generally one of data - more data, more correlations, more accurate predictions. There was a great paper that I read that basically said ‘if you have enough data, the algorithm doesn’t even matter that much’, and empiracally proved it. Now, if I can just find that paper again….
March 29th, 2007 at 7:34 am
Netflix eh? I wouldn’t have guessed, but it makes sense because the consumer eventually logs in again to rent another movie.
Consumer data would definitely make recommendations better. Though, as a lazy programmer, I’m guessing that algorithm includes consumer data (liked or hated) plus some kind of prediction mechanism, which would be based on attributes of the item. In netflix case, it would make sense to include movie genre, actors, and similar details, but do they go further to include plot attributes - sad ending, happy ending, car chases, train rides, type things? It probably would, because that’s what goes through my mind when deciding whether to watch a movie or not.. So can we apply that same logic to everything else?
If I say I like 2 books from indigo, that should be enough to generate at least one good book recommendation, no? Though, getting the “i like” data is easier said than done.