% Similar?

[The Professor]

I rarely see other Netflix members who are more than 70% similar to me. Recently though, I got a friend request that I accepted that came in at a whopping 94%! I'm wondering what sort of similarity percentages others see, and what they think of the whole idea...

[I'm Not a Turnip]

I always see people in the mid 80s at the bottom of the Friends page. Yesterday I made one a Fave because she was listed as 88%. Today she is at 84%. Doesn't seem very accurate to me.

[Slappy]

I have seen it go up and down in minutes so I don't trust it too much.

[2McAbre]

I don't put much into it either, I have one around 93%, but the only same was the movies we rated, their opinions and reviews were not at all helpful to me as they were all about the other 7% we did not have in common. On the other hand I have one that is 26% similar and I enjoy reading that persons reviews and have actually rented a few movies I normally would not have as a result.

So maybe the system is setup backward, look at folks with lower %Sim to see if anything that have rated/ reviewed highly is something you might actually enjoy that you otherwise wouldn't have looked at.

[SH 91980]

I have a couple of 82% similar that I've added, but the bulk are 70's. I also like having contrary views to mine so I get a wide range of movies rather than just the same old stuff I've would pick to watch. If you go through and look at the movie's and how they rated as compared to favs, I am hardpressed to see how I am that similar with many of my favs. I, too, agree that something is off on the "sim ratings". I usually add because I like their writing style or admire their viewpoint, not usually because they are the same. But, come to think of it, I don't think I've ever seen a rating higher than in the 80's.

[Robowriter]

I see similarity % figures jump 10-20% before and after adding a new friend. I asked tech support about this specifically as well as the numbers that fluctuate from minute to minute. It seems to be tied to the distributed network where not all information is available at all times (particularly when Thu network updates are being run) and he surmised that the sim% figures you see on the reviews page are only calculated on a quarterly basis while the sim% figures on your friends page are calculated on an as-loaded basis. I didn't get the impression that he *knew* and had been trained to give this as the answer but that it was his best guess for a cause or reason.

[Mastermindobu]

I may be wrong in my assumption, but I think the reason the percentages change is that you can only see/compare your five star rated movies before you are friends. Once you become friends it takes into consideration all the movies that you've ever rated.

The highest on my friends list is at 91% and it's pretty dead-on. I have someone that is 81% on my account but I am 99% on his account.

[Robowriter]

I'm curious how my sim% rating for friend A can differ from his (or hers) for me.

I'm even more curious how I can rate half the titles in a series as 4 stars and Netflix still (after weeks) predicts that my rating for the rest of the series will be 1 star.

I'm even more curious why the Top 10 Lists screen says at the top "We suggest these lists based on your ratings" and then gives me a list of lists that's 50% wrestling and gay discs. Where are the genres I have generally marked as 5 stars instead of Not Interested?

[droidmaker]

I know i've written this before (somewhere), but i'll give it a shot to the best of my recollection. First of all: i'm the one who came up with this sim% and worked on the algorithm, so i'm not guessing here.

1. There is no perfect way to determine the movie similarity tastes between two individuals. There are lots of "interesting" ways - and some produce good results in one dimension but not another. It is not meant to be "absolute" nor "definitive" - but it should be a reasonable way to differentiate between you and others in terms of the liklihood that they will have some interesting input for you.

2. When originally conceived, the sim% was between friends who explicitly asked to be connected. Since people have few friends, it was a good calculation to run that could be relatively complex (and thus, slower). We used ratings data to compare folks. It was weighted to know that absolute distance between ratings isn't always the same (e.g the difference between a 4 and a 5 star rating is smaller than the difference between a 2 and a 3 or a 3 and a 4. And the difference between a 3 and a 5 is far less than the difference between a 2 and 4. I'm simplifying, but you get the idea. )

3. The sim% are not "symmetrical" because if you have rated 1000 movies and I have rated 100 movies -- all 100 of mine may be a subset of your 1000, so even if we match perfectly on the ratings of those movies, there is a huge discrepancy in the movie experiences we have. This is a matter of utility. The odds that me (with 100 ratings) would have something to share with you (1000 ratings) isn't that great, but the heavier movie watcher has more utility to the light one. We build some sense of utility into the sim% which causes this assymetry. It was a compromise to keep things simple, but i agree sometimes, it feels weird to see a different number on your screen looking at me, than i see of you. I know just what you mean. But i like the idea of utility. Particularly if you sort those rows by Sim%.

4. Strangers. There are too many people in the netflix database to rapidly cacluate the sim% between every pair. It put a drain on the system. In our exploration to see if we could provide utility to members quickly, we came up with a faster version that looks more at genres and subgenres of interest. It doesn't look at ratings at all. But it looks at movies in both your history as well as your queue, and measures overlaps and depth of interest in types of movies. Every movie has a range of genres/subgenres associated with it, so even if YOU think of a movie as "horror", it's possible it has also been classified as "foreign" or "strong female leads" -- if you don't think your profile accurately describes you, well, i understand that. Over the months we created that stuff, we tweaked the various weights of the elements and tuned it as best we could. It never solved all problems or didn't have oddities, but it was remarkably good for a lot of people.

I'm not sure of the granularity of "accuracy" there, but my friends (who i often share movies) are in the high 70s. And only a couple of special folks are in the 80s. I think 80s is a very high sim%. From 50s-60s% I don't think of these as particularly great matches for me. And I find that when it says someone is in the 40s - i agree, i don't have much in common with this person.

Maybe we should have used a simple thumbs up (>60) or thumbsdown (<59) - but the %s felt a little more informative and fun.

Do you think Netflix should remove it?

[Mastermindobu]

Thanks for the info. I like having it; it's definitely not perfect but it does give a good indicator.

[pterosaur]

Thanks, Droid! I've never seen that explanation before.

Accurate or not, I don't think you should remove it.

[Flix_Lover]

I don't think that Netflix should remove it. I find it to be a fun thing. Every now and then, I come across some stranger's profile who's rated about 80% or more similar to me; so, I look at their profile and usually end up adding a couple of movies to my queue based on what they've rated, reviewed, or listed on their top 10 lists.

[BitPart]

WORD.......