Here's the dataset they use. I have used this as part of developing and testing the fitness of recommender systems in the past: http://grouplens.org/datasets/movielens/
This looks interesting, but I'm always sad to see consumer web apps still requiring you to sign up before you can try it out. Missing out on so many potential users. 1) sell the product then 2) ask for the customers info
Agree. If there was a way to try it out without creating an account, I probably would. As it is, there is no chance unless for months I consistently hear from friends "You have to try MovieLens - it's brilliant." And I've never heard about it, so I close the tab.
Movielens is non-commercial product. Its emerged and maintained as a product of Grouplens Research group. They may not have put effort to integrate with some obvious SSO but its definitely worth a try.
P.S. I am biased because I have worked in Grouplens research lab for a short time and I have loved interacting with all the students and professors there.
Yes, I'm aware of the background, but surely if they want people to use it, they should take the best route to achieving that? Or if it's a learning experience, learn part of the reality about getting people to use a service?
Otherwise, how many of us look over the page and then walk away?
Maybe ask the user to enter three movies they like and then give a recommendation based off those three. I'd be more compelled to sign up once I see the recommendation it returns isn't completely off the mark.
That's a lovely idea in theory, but in practice any recommendations made off three data points are far more likely to discourage than encourage you from ever using the service again.
Recommendation engines require a lot more training (by you) to be useful (to you).
Worse yet, without the modest investment in time, including thinking up a new password, there's a very high risk people will randomly click just to see what the results are, which muddies the water for other users.
In any case, non-identified recommendation source data would need to be kept in a silo until the user identified / validated they were not just fiddling around. And that becomes a major mess to manage the data.
Having said that, I do respect people's right to tilt endlessly at windmills in HN comments.
Sure, but the big problem is not so much the tracking (IP, agent, cookie, etc) but the small number of data points.
To properly evaluate the quality of the data you can get out of a recommendation system, you need to put a useful amount of half-way decent data into it.
It's a relatively small cost to create an account -- it took me less than 30s, most of which was recording a new set of credentials in my password store.
One possibly intentional side-effect of requiring even a very simple registration process is that it may reduce the amount of cruft going into the engine. </speculating>
All that is all good and well, but for two things:
- it took me less time to input enough ratings to get useful recommendations than to register
- the registration asks for unnecessary information
Solution: if you don't want cruft in your dataset throw away recommendations not linked to a registration. If the users finds the recommendations useful, offer to save the profile by registering.
It's not like you are going to get any less cruft by forcing users tonregister. The only difference is that you are going to end up with a lot of throwaway accounts.
You could remember their ratings with a cookie and generate recommendations based on that, with a banner on every page saying "Save your ratings, make an account".
Once people have invested a few minutes in rating movies they're probably much more likely to want to save that by making an account.
http://criticker.com/ is a good alternative. For each movie, it shows you ratings from users sorted by their rating-similarity to you. I've found it really helpful to see all the ratings in addition to getting a predicted rating. I often use the individual ratings to figure out whether a movie is in the cluster of what people-like-me watch, and also to see whether opinion is unanimous or divisive.
It might be great, but personally I don't like the UI. The throwback layout, poor spacing and positioning spiced with non-relevant and distracting ads just put me off. But then again that's just me, I bet somebody loves it.
I do however take issue with numerical scoring that is then changed to a value based word score upon submission. People are notoriously bad at scoring things on a 0-100 scale. It also begs the question why there is 0-100 scale when a non-disclosed word score is applied anyway.
The final nail in the coffin was that they lost my activation email and won't let me in without it. To add insult to injury there isn't an option to resend the email. The only way to proceed is to email support and wait upon their indulgence.
> then changed to a value based word score upon submission
Your numerical score is still there. You can pretty much totally ignore the descriptors and think of the tiers T1-T10 as percentiles.
> begs the question why there is 0-100 scale
This lets you rate to any granularity you like to precisely order your movies. After you have a large number of rated movies, you might want to tweak the scores to bump a movie into a higher or lower tier.
I rated 43 films, then selected "Top Picks for You" and got 0 results. So I decided to not watch a movie and instead go for a walk. Which turned out to be exactly what I needed. Five stars.
It seems that recommendations by the "wizard" mode are updated daily or so. I signed up yesterday and got nothing with the wizard mode. It starts to show promising results just now I checked.
Despite there being so many of these sites, most of them kind of fail for me.
For instance... last night I was wanting to watch a movie, but the kids needed to watch it too (as they can't be left unattended with me watching a movie in the other room :)).
So I went off on a search for a movie to watch... with a rating filter -- it needed to be either G or PG.
Nothing. No recommendation site, Netflix, or anything let me filter movies/shows this way. Yes, you have the 'Kids' section on Netflix, but most of the stuff there is deathly boring for adults, I was hoping for something good for both kids and adults.
Although not a personal recommendation site, instantwatcher.com allows you to filter Netflix titles by rating and sort by a lot of other useful criteria such as production year, rotten tomato rating etc.
If you have an iPad I made an app called Dejavu that might be able to help, you can filter by genre and year of release, and if you think it would be good I could potentially add filtering by rating. Check it out :) I'd love to get some feedback
https://itunes.apple.com/app/id914369680
Something like this would be useful if you could cross-reference your recommendations with others. If I have a set of recommended movies, tell me which movies are also recommended to members of my family, so that we can find the one we are most likely to all enjoy.
Shit, I hadn't heard that John Riedl died. I'm really sorry to hear it. I never had a chance to take a class with him, but Professors Konstan and Terveen were incredibly influential for me.
Proof that Netflix doesn't need all that effort to deliver a laggy, annoying, UI. Seriously, the core functionality is essentially the same. MovieLens seems like it might replace Netflix for me. Not-terrible recommendations and a decent UI. Though I've only been using it a few days so we'll see, but it can barely end up worse than Netflix eh?
I'm torn. On the one hand, I just spent a few minutes rating about 100 films, and the quality of the recommendations is so much better than Netflix's garbage that it's mindblowing. As in, we've spent half an hour browsing Netflix trying to find something to watch and settling for something meh, whereas here I have a full page of recs with 5 minutes input.
On the other hand, I don't know if I'm very impressed with anything yet, the recommendations seem obvious. It's more like a neat visual collection manager, I'm not sure I "feel" anything crazy going on in the background beyond what looking at IMDb best-of lists would produce. In other words, the value seems to be in the Netflix-ish "cover collection" display, unconstrained by the limits of their selection.
This predates some of my more recent work with the grouplens database, but here is a parser I put together for the data awhile back: https://github.com/posborne/mlcollection/blob/master/mlcolle...