Item recommendations - work in progress

I've been working on a simple recommendation algorithm for individual items, I didn't expect too much but the results seems ok for a lot of things, to access them you just have to add /recommended to the url e.g. :

www.listal.com/movie/finding-nemo/recommended
www.listal.com/tv/24-2001/recommended
www.listal.com/movie/the-matrix/recommended
www.listal.com/artist/radiohead/recommended
www.listal.com/author/alan-moore/recommended
www.listal.com/book/v-for-vendetta-alan-moore/recommended
www.listal.com/music/ok-computer-radiohead/recommended
www.listal.com/game/halflife-2-pc/recommended

It will work for books, games, music, authors, artists, movies, tv, dvd and directors

As you can see it turns up some pretty randon things but also gets it right a lot of the time as well :)

I would be interested in feedback on these results, I've mainly tested tv/movies/games. My knowledge of music is pretty bad so I'm not sure how good these recommendations are.

The recommendations are not filtered for genre but have the option of doing this in the future.

Looks like a reasonable start. Do like the simple layout of them - nice and uncluttered.

Music recommendations - artists are mostly in the right ballpark of genre from the ones I've looked at, but it's not really choosing the most suitable artists - rather it's choosing the most popular (Sonic Youth, Tori Amos and Radiohead seem to appear in virtually every page I've looked at, whether appropriate or not). This is probably somewhere where last.fm would be good too look at to tune the engine. (compare www.last.fm/music/The+Libertines/+similar to www.listal.com/artist/the+libertines/recommended as results for the same band). I certainly think that someone that likes the particular band would find last.fm's matches more suitable than yours at present.

Releases on the other hand, look fairly good on the whole from the 10 or so I've looked at. Main suggestion there would be to increase any factor you've got on releases by the artist that released that particular CD/vinyl. www.listal.com/music/nevermind-nirvana/recommended for example, I'd have thought In Utero would be if not the top item, second at worst. Similarly with www.listal.com/music/weezer-weezer/recommended, most of Weezer's back catalogue doesn't feature, despite being a very clear progression on their debut (this album!).

Not knowing what factors you're using, if the formula was something on the lines of x*is_this_by_same_artist + y*how_many_owners_have_both + z*how_much_owners_of_both_like_it, then the 'x' factor would want increasing. Once artist results are improved, then you could improve these results even more by factoring artist similarity in rather than a simple "is this by the same person(s)' check.

Hope that makes sense and helps!

What about using the most popular tags as one of the criteria for filtering recommendations?

Taking Finding Nemo as example, Saw appears on the recommended list but when you compare the top tags of each movie, they don't have anything in common.
Finding Nemo: animation, family, comedy, pixar
versus
Saw: horror, murder, kidnapping, thriller

I thought about integrating tags and excluding unrelated items, I will experiment with that. The recommendations system I have created will return items liked by people who like the item you are viewing and is not designed to recommend similar items especially (although these often show up).

I would be interested in how last.fm generate their similar artist.

It's based on tags, listening data and some artist data that they've got (gleaned I believe from Musicbrainz) as well as black magic. The latter, for example, would show that Graham Coxon was a member of Blur so he would be more likely to be recommended than an artist that isn't. Tags is fairly obvious - as Vix mentioned, it tweaks the results somewhat to try and ensure that entirely dissimilar artists aren't recommended (death metal for a classical composer.

The listening data is simply (people who listen to this, also listen to this). However, there's a rarity factor involved as well to ensure it's not dominated by the top few artists on the site. I guess the best way to look at it would be on the lines of artist A has 1,000,000 listeners, while artist B only has 1,000. However, of those 1,000, 90% also listen to artist A, so it gets a high rating despite only 0.9% of listeners to artist A also listening to artist B. It gives lesser known artists/albums/films/novels etc a chance.

Obviously, in that algorithm, there has to be a cut off on how many individuals do use both - if it's just one or two, then it could be unduly influenced. The results do seem to be pretty good though.

You might also be interested in using their feeds at www.audioscrobbler.net/data/webservices/ - looks like you'd be covered by their free commons licences at least for the time being, and it'd do a good job of augmenting the data that's still being built up here.