Simplest search for similar movies would be to select from the list of the tags (one that existing movie has) several by which I want to filter similar movies. Then system would show only movies with relevant tags, which would make searching a lot easier.

Half of that is done by the movie tuner. The developers just need to make a little more obvious that it is tag based and how it works. And make it awailable right from the movie details page, not two unrelated clicks away.

As far as similarities go, there'd be 'binary' similarity vs. 'interpretive' similarity. It would be relatively trivial, I'd think, to scrape online data for 'binary' categories like name of a director, actor, script writer, etc. and limit the results to matches. Hence, binary. Either the movies have the same director or they don't. The 'interpretive' similarities need more thought - and the results would be more like the visualization technique over here: http://www.movie-map.com/brazil.html ... for example, if you could scrape online data and add the movie's setting to the database, then the closest movies would be those set in the same place, the next closest would be those set in the next level of the setting heirarchy, i.e. Paris, France, Europe...

The scatterplot with sliders is what I think of as Shneiderman's starplot. But it is the best UI I've thought of so far.

I often look for movies to match my mood for example something like Lubitsch's Heaven Can Wait (1943) but that I haven't seen yet. There are currently tags for that movie that help me find similar movies ("the comfort food of movies", "avuncular grandfather figure", and "autobiographical view"). The correlation view won't work as precisely as something using those tags.

Max, if you want to we could brainstorm a bit by Skype.

Hi again,
just getting back to the tagging issue: I think that if users find the results helpful then they will contribute to the process that makes those results possible. That reasoning works for the recommendations that come from ratings too. Users who aren't sufficiently interested in finding similar movies won't bother with the tagging.

The current interface has a very simple version of this: using item rating correlations, show the 10 most similar movies, ranked by correlation score.

So, separating out the computation piece and the visualization piece:

* There are lots of ways of computing "generic" similarity scores. We have two that we've used in the past that often work - rating correlations and tag similarity. I thought you were getting at "specific" similarity measures, where we look at particular dimensions of a movie to find different sorts of relationships, which is an idea I like.

* The easiest way to visualize similar movies is, starting from an anchor movie (e.g., on a movie page), show a list of related movies. I am open to a more complicated navigation interface (e.g., a scatterplot with controls) if I can understand the types of browsing and exploration that it will enable. That is not clear to me yet, though I'm happy to brainstorm.

Hi, Max.

Perhaps it is my paucity of imagination but I was thinking of graphs (the math kind) perhaps with clusters. I think that anything like Shneiderman's star field would be overkill.

As to how to make it happen I think that tags are the best way right now; but it would require some way to combine tags (for example, Paris and setting:Paris). It might not be too much to hope that if users see an advantage in these types of tags then more tags will appear and the approach will be feasible.

Thinking more about how we could get this feature. As a user, I can offer to say which movies are similar to the movie I'm currently looking at. The system could prompt me for movies that are similar in particular ways (setting, remakes, and series come to mind), and display these relationships separately in the interface. Interesting possibilities here.

Our tagging interface is flexible enough to support some of these ways of organizing movies (e.g., setting), but we do not have any current mechanism for organizing users into "tagging campaigns" that would lead to concerted tagging effort. And, without concerted effort, maybe 1% of the database gets tags related to setting.

Do you have other ideas for how to collect this data about movies?