IMDb top 250 list with PR
I've long thought that the IMDb top 250 list https://www.imdb.com/search/title/?groups=top_250&sort=user_rating would be more interesting if it was done using a sequential PR method. That way you would likely get more of people's absolute favourite films rather than those with broad appeal, but which might not be right at the top of people's lists. You're more likely to see differents genres represented etc.
Anyway, not that it will do anything, but I Tweeted at them today with my suggestion here: https://twitter.com/toby_pereira/status/1624513669332717569
By the way, I definitely think that Sequential proportional score voting https://electowiki.org/wiki/Sequential_proportional_score_voting (also known as SPAV + KP) is the best method for this. When you're electing an essentially unlimited number of items in a sequential manner, you do not want to be messing about with quotas etc. A Thiele method is ideal for this, and as it's for scores (well stars), I consider this to be the best version as it passes multiplicative and additive scale invariance, which other methods do not. IMDb uses a 1 to 10 star scale, and if 1 was subtracted from every score to make it 0 to 9, then the results under Sequential proportional score voting would be identical.
I don't generally use Twitter, but feel free to engage with the Tweet to like retweet etc. so it can gain more traction (if you agree with it obviously).
I know that this is a more "trivial" matter than is normally discussed, but I see it as a demonstration that electoral methods have uses outside of elections for public office.
Toby
On 11.02.2023 22:27, Toby Pereira wrote:
I've long thought that the IMDb top 250 list
https://www.imdb.com/search/title/?groups=top_250&sort=user_rating would
be more interesting if it was done using a sequential PR method. That
way you would likely get more of people's absolute favourite films
rather than those with broad appeal, but which might not be right at the
top of people's lists. You're more likely to see differents genres
represented etc.
I think I wrote a post about more or less this use of proportional
representation, but I just can't seem to find it. I said that I was
looking for a method that would work on very large numbers of candidates
and voters, and a reply suggested the use of SPAV/SPRV.
In my case, I was thinking about recommender systems (which is another
subject of interest of mine); the simplest recommender systems take the
current user's preferences (watched movies/ratings) and determines a
similarity to other users', then does a weighted sum of those
preferences to create suggestions for the users in question.
In essence, the other users' preferences become weighted rated or
Approval votes (depending on whether the system gathers ratings or just
thumbs up/down). There are other tricky parts when the recommendations
are implicit (e.g. they consist of movies the user has watched rather
than explicit thumbs up/down) because it's much harder to determine
whether a user didn't watch because he didn't know about the movie, or
because he knew he wouldn't like it.
In any case, summing the ratings predictions by other users becomes a
kind of approval/range voting process. But the output is often "hey, you
liked Casino Royale; here I will suggest every other Bond movie". So
proportional representation makes a lot of sense. But since the
database could potentially be extremely large (both in terms of
users/voters and items/candidates), STV can't work, and methods that
aren't polytime (e.g. Schulze STV) are right out. So SPAV/SPRV does seem
the right type of voting method to use in such a context.
I implemented SPAV on a basic recommender system for another site, and I
think it provided more varied answers, but because "users who liked what
you liked also liked" type reasoning can't get more subtle relations,
the quality is still limited: I got a sort of "cycling behavior". E.g.
suppose a user had watched Casino Royale and Interstellar, SPRV/SPAV
would then produce something like [Spy movie] [Spy movie] [Spy movie]
[Sci-fi] [Sci-fi] [Sci-fi] [Spy movie] [Spy movie] [Sci-fi] [Sci-fi]...
Which is better than just all spy movies all the time, but could still
be improved upon :-)
Current recommender systems are considerably more advanced than the
nearest-neighbor type above, but I think they miss an active learning
component, the lack of which leads them to be either too uninteresting
(recommending the same thing over and over), or have too little
information to say anything at all (thus recommending mainly things that
appeals to everybody).
Xing et al.'s paper, "Enhancing collaborative filtering music
recommendation by balancing exploration and exploitation",
https://archives.ismir.net/ismir2014/paper/000140.pdf, gives one way to
do this. They focus on the problem that recommending the same thing over
and over leads to the user getting bored and suggest a Bayesian approach.
Another approach could be having an interactive system where the user
can choose between different groups or clusters of items depending on
his interest; and when one group is chosen, the system then produces
subgroups within that group. E.g. first having a choice between spy
movies and sci-fi, and then Carré-type vs Mission Impossible-type spy
movies, or between space fantasy and near-future sci-fi (e.g. Star Wars
vs Gravity).
It might even be possible to combine Bayesian low rank matrix completion
with Thompson sampling to get a more comprehensive sort of active
learning. But I'm not sure if BLRMC is just a neat trick to make low
rank matrix completion (relevant for latent factor type recommenders)
into a maximum posteriori estimation problem, or if the distribution
itself can be used. And it might be much too hard to sample anyway.
Whew, this got a bit off-topic. But I guess I'm saying... recommender
systems are not a trivial matter either :-)
-km
The best kept secret in electoral reform is that the General Medical
Council, in Britain, used to elect all white male GPs First Past The
Post. Until in 1979, STV proportionally represented women, immigrants
and specialists, in the National Health Service. (The Best System, by
The Electoral Reform Society, 1984. Centenary celebration. archive.org )
Hollywood (or whoever) who I believe use RCV (ranked choice voting)
could apply an At-large RCV proportional count (alias STV/PR). And its
conventional film categories would "show" likewise proportionally
represented.
STV would be better than a points system, from the scientific viewpoint
of the scales of measurement. A points system implies an interval scale
(with no true zero, like a temperature scale, such as Centigrade). STV
implies a more powerful ratio scale. With traditional STV the ratio
scale only applies to the election count but experience has shown this
to be sufficiently robust. And there is not much difference between STV
versions in the results.
Nevertheless, Traditional STV only uses an ordinal scale "last past the
post" kind exclusion count. Binomial STV uses both a rational election
count and a rational exclusion count, which is just an iteration, with
the preferences reversed. However, the fact that Binomial STV gives
voters the power to exclude, as well as elect candidates, means that it
should give significantly different, perhaps more comprehensive
assessments of positive and negative voter attitudes.
Regards,
Richard Lung.
On 11/02/2023 21:27, Toby Pereira wrote:
I've long thought that the IMDb top 250 list
https://www.imdb.com/search/title/?groups=top_250&sort=user_rating
would be more interesting if it was done using a sequential PR method.
That way you would likely get more of people's absolute favourite
films rather than those with broad appeal, but which might not be
right at the top of people's lists. You're more likely to see
differents genres represented etc.
Anyway, not that it will do anything, but I Tweeted at them today with
my suggestion here:
https://twitter.com/toby_pereira/status/1624513669332717569
By the way, I definitely think that Sequential proportional score
voting
https://electowiki.org/wiki/Sequential_proportional_score_voting (also
known as SPAV + KP) is the best method for this. When you're electing
an essentially unlimited number of items in a sequential manner, you
do not want to be messing about with quotas etc. A Thiele method is
ideal for this, and as it's for scores (well stars), I consider this
to be the best version as it passes multiplicative and additive scale
invariance, which other methods do not. IMDb uses a 1 to 10 star
scale, and if 1 was subtracted from every score to make it 0 to 9,
then the results under Sequential proportional score voting would be
identical.
I don't generally use Twitter, but feel free to engage with the Tweet
to like retweet etc. so it can gain more traction (if you agree with
it obviously).
I know that this is a more "trivial" matter than is normally
discussed, but I see it as a demonstration that electoral methods have
uses outside of elections for public office.
Toby
Election-Methods mailing list - seehttps://electorama.com/em for list info
Thanks for the insights, Kristofer. In a way, recommender systems (as I generally understand them) are the opposite of a proportional list, as recommender systems tend to end up with lots of things that are similar. But it's still conceptually similar and you might want a recommender system to suggest a more diverse range of stuff.
Toby
On Monday, 13 February 2023 at 13:25:11 GMT, Kristofer Munsterhjelm km_elmet@t-online.de wrote:
On 11.02.2023 22:27, Toby Pereira wrote:
I've long thought that the IMDb top 250 list
https://www.imdb.com/search/title/?groups=top_250&sort=user_rating would
be more interesting if it was done using a sequential PR method. That
way you would likely get more of people's absolute favourite films
rather than those with broad appeal, but which might not be right at the
top of people's lists. You're more likely to see differents genres
represented etc.
I think I wrote a post about more or less this use of proportional
representation, but I just can't seem to find it. I said that I was
looking for a method that would work on very large numbers of candidates
and voters, and a reply suggested the use of SPAV/SPRV.
In my case, I was thinking about recommender systems (which is another
subject of interest of mine); the simplest recommender systems take the
current user's preferences (watched movies/ratings) and determines a
similarity to other users', then does a weighted sum of those
preferences to create suggestions for the users in question.
In essence, the other users' preferences become weighted rated or
Approval votes (depending on whether the system gathers ratings or just
thumbs up/down). There are other tricky parts when the recommendations
are implicit (e.g. they consist of movies the user has watched rather
than explicit thumbs up/down) because it's much harder to determine
whether a user didn't watch because he didn't know about the movie, or
because he knew he wouldn't like it.
In any case, summing the ratings predictions by other users becomes a
kind of approval/range voting process. But the output is often "hey, you
liked Casino Royale; here I will suggest every other Bond movie". So
proportional representation makes a lot of sense. But since the
database could potentially be extremely large (both in terms of
users/voters and items/candidates), STV can't work, and methods that
aren't polytime (e.g. Schulze STV) are right out. So SPAV/SPRV does seem
the right type of voting method to use in such a context.
I implemented SPAV on a basic recommender system for another site, and I
think it provided more varied answers, but because "users who liked what
you liked also liked" type reasoning can't get more subtle relations,
the quality is still limited: I got a sort of "cycling behavior". E.g.
suppose a user had watched Casino Royale and Interstellar, SPRV/SPAV
would then produce something like [Spy movie] [Spy movie] [Spy movie]
[Sci-fi] [Sci-fi] [Sci-fi] [Spy movie] [Spy movie] [Sci-fi] [Sci-fi]...
Which is better than just all spy movies all the time, but could still
be improved upon :-)
Current recommender systems are considerably more advanced than the
nearest-neighbor type above, but I think they miss an active learning
component, the lack of which leads them to be either too uninteresting
(recommending the same thing over and over), or have too little
information to say anything at all (thus recommending mainly things that
appeals to everybody).
Xing et al.'s paper, "Enhancing collaborative filtering music
recommendation by balancing exploration and exploitation",
https://archives.ismir.net/ismir2014/paper/000140.pdf, gives one way to
do this. They focus on the problem that recommending the same thing over
and over leads to the user getting bored and suggest a Bayesian approach.
Another approach could be having an interactive system where the user
can choose between different groups or clusters of items depending on
his interest; and when one group is chosen, the system then produces
subgroups within that group. E.g. first having a choice between spy
movies and sci-fi, and then Carré-type vs Mission Impossible-type spy
movies, or between space fantasy and near-future sci-fi (e.g. Star Wars
vs Gravity).
It might even be possible to combine Bayesian low rank matrix completion
with Thompson sampling to get a more comprehensive sort of active
learning. But I'm not sure if BLRMC is just a neat trick to make low
rank matrix completion (relevant for latent factor type recommenders)
into a maximum posteriori estimation problem, or if the distribution
itself can be used. And it might be much too hard to sample anyway.
Whew, this got a bit off-topic. But I guess I'm saying... recommender
systems are not a trivial matter either :-)
-km
On 2/17/23 14:01, Toby Pereira wrote:
Thanks for the insights, Kristofer. In a way, recommender systems (as I
generally understand them) are the opposite of a proportional list, as
recommender systems tend to end up with lots of things that are similar.
But it's still conceptually similar and you might want a recommender
system to suggest a more diverse range of stuff.
Yeah; everybody who's had the "you bought a kitchen stove, here are
fifteen more kitchen stoves you may be interested in" effect happen to
them would know that diversity is a benefit :-)
I guess the link is that recommenders are diverse across different
people. If you can't customize your response to the recipient at all,
then the only way to be diverse across different people is to do
something like PR or consensus voting.
... and that what I'm saying is just that they should also be diverse
over time, when they don't know who the person is or if his tastes have
changed. In a way, the same person at different times become different
people.
(Consensus voting is my term for things like minmax approval, where the
outcome is so that it's liked most by the voter who likes it the least.
E.g. if you're buying pizzas for a party, everybody should be able to
choose at least one pizza they like. I don't know what the formal term
for these method types are.)
-km