<html><head></head><body><div class="ydp811c2221yahoo-style-wrap" style="font-family:Helvetica Neue, Helvetica, Arial, sans-serif;font-size:16px;"><div></div>
<div>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.</div><div><br></div><div>Toby</div><div><br></div>
<div id="ydp811c2221yahoo_quoted_7029545097" class="ydp811c2221yahoo_quoted">
<div style="font-family:'Helvetica Neue', Helvetica, Arial, sans-serif;font-size:13px;color:#26282a;">
<div>
On Monday, 13 February 2023 at 13:25:11 GMT, Kristofer Munsterhjelm <km_elmet@t-online.de> wrote:
</div>
<div><br></div>
<div><br></div>
<div><div dir="ltr">On 11.02.2023 22:27, Toby Pereira wrote:<div class="ydp811c2221yqt2688719724" id="ydp811c2221yqtfd50437"><br clear="none">> I've long thought that the IMDb top 250 list <br clear="none">> <a shape="rect" href="https://www.imdb.com/search/title/?groups=top_250&sort=user_rating " rel="nofollow" target="_blank">https://www.imdb.com/search/title/?groups=top_250&sort=user_rating </a>would <br clear="none">> be more interesting if it was done using a sequential PR method. That <br clear="none">> way you would likely get more of people's absolute favourite films <br clear="none">> rather than those with broad appeal, but which might not be right at the <br clear="none">> top of people's lists. You're more likely to see differents genres <br clear="none">> represented etc.</div><br clear="none"><br clear="none">I think I wrote a post about more or less this use of proportional <br clear="none">representation, but I just can't seem to find it. I said that I was <br clear="none">looking for a method that would work on very large numbers of candidates <br clear="none">and voters, and a reply suggested the use of SPAV/SPRV.<br clear="none"><br clear="none">In my case, I was thinking about recommender systems (which is another <br clear="none">subject of interest of mine); the simplest recommender systems take the <br clear="none">current user's preferences (watched movies/ratings) and determines a <br clear="none">similarity to other users', then does a weighted sum of those <br clear="none">preferences to create suggestions for the users in question.<br clear="none"><br clear="none">In essence, the other users' preferences become weighted rated or <br clear="none">Approval votes (depending on whether the system gathers ratings or just <br clear="none">thumbs up/down). There are other tricky parts when the recommendations <br clear="none">are implicit (e.g. they consist of movies the user has watched rather <br clear="none">than explicit thumbs up/down) because it's much harder to determine <br clear="none">whether a user didn't watch because he didn't know about the movie, or <br clear="none">because he knew he wouldn't like it.<br clear="none"><br clear="none">In any case, summing the ratings predictions by other users becomes a <br clear="none">kind of approval/range voting process. But the output is often "hey, you <br clear="none">liked Casino Royale; here I will suggest every other Bond movie". So <br clear="none">proportional representation makes a lot of sense. But since the <br clear="none">database could potentially be extremely large (both in terms of <br clear="none">users/voters and items/candidates), STV can't work, and methods that <br clear="none">aren't polytime (e.g. Schulze STV) are right out. So SPAV/SPRV does seem <br clear="none">the right type of voting method to use in such a context.<br clear="none"><br clear="none">I implemented SPAV on a basic recommender system for another site, and I <br clear="none">think it provided more varied answers, but because "users who liked what <br clear="none">you liked also liked" type reasoning can't get more subtle relations, <br clear="none">the quality is still limited: I got a sort of "cycling behavior". E.g. <br clear="none">suppose a user had watched Casino Royale and Interstellar, SPRV/SPAV <br clear="none">would then produce something like [Spy movie] [Spy movie] [Spy movie] <br clear="none">[Sci-fi] [Sci-fi] [Sci-fi] [Spy movie] [Spy movie] [Sci-fi] [Sci-fi]...<br clear="none"><br clear="none">Which is better than just all spy movies all the time, but could still <br clear="none">be improved upon :-)<br clear="none"><br clear="none">Current recommender systems are considerably more advanced than the <br clear="none">nearest-neighbor type above, but I think they miss an active learning <br clear="none">component, the lack of which leads them to be either too uninteresting <br clear="none">(recommending the same thing over and over), or have too little <br clear="none">information to say anything at all (thus recommending mainly things that <br clear="none">appeals to everybody).<br clear="none"><br clear="none">Xing et al.'s paper, "Enhancing collaborative filtering music <br clear="none">recommendation by balancing exploration and exploitation", <br clear="none"><a shape="rect" href="https://archives.ismir.net/ismir2014/paper/000140.pdf, " rel="nofollow" target="_blank">https://archives.ismir.net/ismir2014/paper/000140.pdf, </a>gives one way to <br clear="none">do this. They focus on the problem that recommending the same thing over <br clear="none">and over leads to the user getting bored and suggest a Bayesian approach.<br clear="none"><br clear="none">Another approach could be having an interactive system where the user <br clear="none">can choose between different groups or clusters of items depending on <br clear="none">his interest; and when one group is chosen, the system then produces <br clear="none">subgroups within that group. E.g. first having a choice between spy <br clear="none">movies and sci-fi, and then Carré-type vs Mission Impossible-type spy <br clear="none">movies, or between space fantasy and near-future sci-fi (e.g. Star Wars <br clear="none">vs Gravity).<br clear="none"><br clear="none">It might even be possible to combine Bayesian low rank matrix completion <br clear="none">with Thompson sampling to get a more comprehensive sort of active <br clear="none">learning. But I'm not sure if BLRMC is just a neat trick to make low <br clear="none">rank matrix completion (relevant for latent factor type recommenders) <br clear="none">into a maximum posteriori estimation problem, or if the distribution <br clear="none">itself can be used. And it might be much too hard to sample anyway.<br clear="none"><br clear="none">Whew, this got a bit off-topic. But I guess I'm saying... recommender <br clear="none">systems are not a trivial matter either :-)<br clear="none"><br clear="none">-km<div class="ydp811c2221yqt2688719724" id="ydp811c2221yqtfd11271"><br clear="none"></div></div></div>
</div>
</div></div></body></html>