Use my viewing habits to sort suggestions, not just predicted ratings.
MovieLens is great at predicting what I might rate films; I've got no problem with that. What it doesn't seem to predict all that well, however, is that I'm rarely going to be in the mood to watch foreign films from before 1970. Yeah, I will probably rate them highly, but some consideration should be made for my actual viewing habits - because those aren't the types of films I've told MovieLens that I regularly watch.
Out of the 542 films that I've rated on MovieLens, only 19 were from before 1970, and only 26 were foreign language. With that in mind:
42 out of my top 50 suggested films are from before 1970. 26 were from before 1960.
39 out of my top 50 suggested films are foreign language films.
It knows I'm able to appreciate certain films, but it doesn't appreciate what films I want to watch on a regular basis. I can deal with this myself via Saved Searches, but if you're going to turn this site into a popular app, you'll definitely want to nip this in the bud. Use people's viewing habits (years, genres, avg ratings) to sort their suggestions, don't just use their own predicted ratings. I know what 5 star films are, but I don't only want to watch them.
Anonymous User commented
I agree with the user below who said "I much prefer the search filters to any sort of "viewing habits" filter - I like deciding what I'm in the mood to watch by seeing all the options. I second the comment below - make it optional."
Viewing habits that filter for me without me agreeing to said filtering are why I hate netflix's interface (and b/c they keep changing the stupid cover art every few minutes so I can't tell if I'm looking at the same title or not).
Also folks talking about no longer being into a certain phase or not being in the mood for something they'll likely really enjoy...I don't get. It's counter-intuitive to me because the reason I am interested in a site like movielens is specifically to find more movies I'll really love (and to think about what makes me love a movie, specifically) rather than the same not great stuff I might end up watching b/c streaming sites push their content on me in a way they think I'll want. Like, it's a GREAT thing when movie lens reminds me how much I like a genre I haven't been watching lately so I can get back into it.
I found movielens because of Jinni. Jinni closed but I really relied on it! I loved all the features and discovered films I didn’t know about. ML is the best I’ve seen to even compare so far. I agree with this post 100%. It’s important to have more fine tuning. We have already watched movies we loved, and want to discover another gem.
1. 5 stars just isn’t broad enough
2. Seeing watch list of other members who also liked my favorites is very helpful, if they want to share.
3. Filters might help
I started with Netflix in the beginning, and I can’t believe how snobby it is now, like “you’ll watch what NF says you’ll watch!” I canceled NF.
A combination of both would be best for me. If I watch 90% blockbusters and 10% arthouse, the recommendations should be 90% blockbusters (the ones with the hightest predicted rating) and 10% arthouse as well.
The grouping could obviously be different and more complex, including characteristics such as language, year of production, genre, country, director, etc.
Mel Lenz commented
I much prefer the search filters to any sort of "viewing habits" filter - I like deciding what I'm in the mood to watch by seeing all the options. I second the comment below - make it optional.
The unique value of Movielens--other than its deadly accuracy--is that it will recommend movies I've never heard of. I can go on lots of other movie sites and get the latest blockbuster brain candy recommendation.
The movieexplorer currently being tested might address this since you can select movies you are in the mood for and not pick that artsy film you loved but don't want to watch at a party or to relax to help guide it.
I would like to add a counterpoint to the suggestion that recommendations be based on viewing habits. It should be optional. That's how they broke Netflix. I would prefer that the recommendations be based purely on my ratings, because my ratings reflect how much I actually liked something, I don't try to be objective about it. Also I sometimes give a negative review to something I haven't seen just because I hate it on principle, such as anything with zombies in it, as a way of saying stop recommending zombie stuff, but if the recommendations are based on what I've rated instead of how I rated it, it could have the opposite effect. If the recommendations are based only on how much you like something then it's easy to get good results just by being honest about how much you like things.
I agree with the op that rating prediction and recommendation should be separated more clearly, e.g. allow me to sort by recommendation or by predicted rating. I think some kind of mood setting for the recommendations would be great indeed. Generally, right now the recommendations seem to mirror too closely the predicted ratings. But I'm not always in the mood for arthouse/foreign movies...
I agree! An easy way to do this would be to integrate with letterboxd, which keeps a diary of your movies -- that info would be transferred across
Kyle Robichaud commented
Ya, this would be great. It would help filter out all the kid's movies out of my recommended list.
This would be super-sweet if you could pull it off. I almost never watch Documentaries so it would be great if Movielens could recognize that and get movies like Man on Wire out of my recommendations list.
I know that this involves a major change, but I think the predictive algorithms would benefit if they knew WHEN did we saw the movies. This is directly related to the original poster of the idea. Some people go through phases, maybe I liked horror movies 10 years ago, now I don't, so don't recommend me horror movies. The time resolution does not have to be accurate (I think. Maybe just the year/month when we saw the movie, or just year, or decade if I don't rememeber exactly - better than nothing)
I should also say that we also want to do a better job of recommendation (not just prediction), and are open to all sorts of ideas like the ones posted here. Keep them coming.
That is a cool idea. We currently compute a data structure that we call the "tag genome", that assigns movie-tag association scores even in the case when a movie does not have a particular tag. You could imagine setting a mood tag for your browsing session, then only displaying movies that meet some threshold association with that tag.
I think 'moods' is a giant problem. Not a crucial one, but a problem nonetheless. Saved searches (with advanced functionality) were some solution to some people.
Possibly a better solution would be some feature dedicated to moods, like lednerg suggests.
Maybe a "how often do you want to watch movies like this" question when rating a movie. Or the option to classify movies out of say 10 (infinite would be better) user-nameable 'mood slots'. On the other hand, global moods doesn't seem too bad. Surely there is general agreement over 'heavy' vs 'light' movies at least. Duplicate tags function, call them 'mood tags'?
I don't really know the solution, I just think the problem is worth solving :)
Thanks for the feedback. We are certainly considering re-adding the saved search feature. We are also considering a "sticky" filters approach, where you can set a "nothing before 1990" filter, and it stays on until you close it. Thoughts?
MovieLens does not currently recognize patterns like those that you mention ("I don't usually like old movies"). We would like to get better at this, but that is a long-term, ongoing effort.