Which content-recommending algorithms are actually good?
For the end-user, that is. I'm sure Meta and the like think their algorithms are fantastic at what they want them to do.
I find myself routinely asking why I get so many suggestions I have no interest in when using all types of websites. I haven't used social media since the early years of Facebook, but I imagine most recommendation algorithms are tuned much like the ones on those sites, i.e. to offer more of the same, whereas I'd prefer something to introduce different stuff I'd probably like. Maybe that differentiates me from the average user, but there should be enough people like me that it'd be factored in, no? Just because I watch a cat video doesn't mean I'm all in on cats.
I mostly like Pandora's service but it feels like their music library isn't huge for my fav genres. Steam regularly tries to interest me in the most insipid games based on superficial commonalities to what's already in my library. Youtube can be good, but it can easily be echo-chambery. Shopping websites of all sorts are usually a crapshoot. What gives?
They're not made for you. They're made to sell product.
So my short answer is: None of them are. I believe that if you want good social media feeds, you have to curate them on your own.
There is a middle ground. YouTube is a great example of this. The content recommendation algorithm on YouTube is bonkers good — it has surfaced absolutely fantastic videos and authors, and I trust mine beyond any doubt.
But: it’s a tool, and it needs maintenance. If I watch a shitty video, I don’t just watch all of it and move on; I dislike it and remove it from my watch history. I actively remove bad recommendations. Etc.
People who complain about the YouTube algorithm being “bad” simply don’t do these things. I’ve even seen people disable their watch history and then complain that the algorithm is bad… well no shit.
I actively curate on YouTube as well, and it still throws me all sorts of random things. It also seems to barely weigh likes and give more weight to "watched and not rated" as whenever I watch a video in a relatively new area and do nothing with it my recommended feed will be flooded with anything tangibly related to that video. Where if I had liked that video that likely wouldn't have happened to the same degree.
Personally, I'd say it is serviceable, not bonkers good.
Yes absolutely, likes are less taken into account than watches. Dislikes are a strong signal however.
YouTube creates cohorts of different types of videos, and then establishes things like “people who watch these types of videos also watch these types of videos”. The suggestions are therefore a mix of “Here are more videos from the same creator”, “Here are more videos like this one”, and “Here are different videos that those who watched this one also watch”.
The homepage is a separate beast, taking the mix of your cohorts into account plus your subscriptions (and how faithful you are to these).
in that world, likes and dislikes have an impact mainly if the video wasn’t previously recommended. A like on a watched video is just an extra “Yeah I didn’t just watch, I enjoyed it” (and is attenuated by how often you like); whereas a dislike is a conflicting signal and if it was on a recommendations it’s an explicit input to reevaluate.
Of course keep in mind all of this is black boxed behind AI models so YMMV. And there are types of videos which this is much more optimised for. Longer content fares better because the engagement signal is much stronger on YouTube; whereas eg TikTok has optimised the other end of the spectrum to give as much signal as possible based on shorter content. So depending what you’re into, it might be bad content for YouTube as well.
I use youtube through invidious because i hate all the crap youtube has, but i believe you. I never likes the recommendation system anyway.
You need to remove it from your watch history. The watch history has way more impact.
Maybe the Roku app is presented with wildly different recommendations, because it's just all kinds of awful.
The suggestions themselves are decent, maybe 50% from existing subscriptions, 25% I'd have interest in, and 25% creators that must be sponsored or something because they're always shoved in when I've never watched them.
However, almost every recommendation row is the exact same thing. Often even in the same order. So there's maybe 20 recommended videos in the entire home page (not including rows for stuff like music and "free with ads" movies). Those recommendations are just repeated endlessly.
One thing I've noticed and come to appreciate about YT's recommendation is that it seems to incorporate what device your account is watching it on in its recommendations. I'm signed into it on the household TV, so my wife watches a lot of videos she prefers there, whereas my desktop browser is very much content I am interested in but she is not. The YT algorithm feeds the two remarkably different homepages and recommendations. It's like it knows what "I" prefer to watch on my TV versus my browser (and, to some extent, versus my phone).
Spotify has pretty much nailed my music tastes, to the point where if I hear any song that I haven't heard before, there's a 70+% chance I'll like it enough to plop it into one of my playlists (or at least my liked songs)
Their podcast recommendations though, jeeze louise! I'm SO easy to recommend to - ALL I listen to are true crime podcasts (and whatever Keith Morrison is hosting cause I could listen to that man read the phone book)!!. But what am I recommended by Spotify?? Weird freaking Dutch influencer podcasts about AI, a podcast by two well known Dutch singers, random people from the Office US show who are doing podcasts about the Office US show??? I might be Dutch but I've literally never listened to a Dutch podcast in my life, and I couldn't give a rats arse about the Office US. Spotify, in the words of my favorite True Crime podcast 'Small Town Murders': SHUT UP AND GIVE ME MURDER! /rant
My guilty pleasure station on Pandora just tossed me this gem a week ago, and I listened to it so much in the first two days I had it completely memorized. It's kind of a lot, but I have a hunch it might be right up your alley.
Penelope Scott - Lotta True Crime
Ha, I love Rat by the same artist
The only annoying thing about Spotify is that there’s no way to indicate you’re not interested in a song/category.
I live in a multilingual country and keep receiving recommendations (like half the auto generated playlists) from a particular language despite not having that language selected.
And they don’t take ‘hide song’ into consideration for the recommendation algorithm!
If you're listening to DJ, or a playlist you own with "smart shuffle" recommendations on, you should see a minus button next to the plus button, at least on desktop.
How well that affects their recommendations is up for debate, but it's something.
Everyone I know who uses TikTok says it has the best algorithm of any social media. They say it feeds the exact content that the person would be interested in.
Ironically I haven’t used TikTok precisely because of these reports. I just don’t want to spend my time doing that.
I tried TikTok just out of curiosity and the people that say that likely are based in the US and don't realize how much of a bubble they are in. I am based in the Netherlands, and I got overwhelmingly Dutch content I have zero interest in. Even after trying for a while it still was complete dog shit to me.
Idk, a lot of my friends here in Russia love TikTok's algorithm, and it shows both Russian and English videos if that's what the person is interested in
That is the thing probably, I have zero interest in 99.9% of the Dutch content I came across while I specifically subscribed and liked English content.
My experience is that TikTok slowly degrades over time. I have to start fresh accounts to “reset” it. Right now all of my videos seem to be split screens including a mobile game and I hate it.
They actually added a "reset recommendations" button into settings a while ago
I wasn’t aware! I’ll try that. I think the issue is they they confuse any engagement with interest.
I will have a day occasionally where my Tiktok is sending me garbage. Some dislikes usually does it, as does a dip into my Following page for a while.
If I have any complaint it's that many of my followed creators get lost even in the following tab.
I have a similar experience with IG but for both if I don’t want to see if I explicitly tell it Im not interested. IG ads are scary good for me now.
Steam actually has a (comparatively) amazing recommendation tool, but it's not heavily promoted: the Steam Interactive Recommender. On default settings, it behaves like a typical recommendation algorithm, but you can tweak it to recommend highly niche or older games. (I don't know for sure what "niche" actually means here; I figure it probably reduces the weight of "general popularity" in the calculation, tailoring recommendations toward stuff that you specifically might like.)
In absolute terms, it's not much, but as someone who is interested in weird ideas and not so much bland polished stuff, the "niche" option in particular is very useful for me.
Wow, I entirely forgot that was a thing. It is indeed very useful.
Neat! I was not aware of that tool at all. Just playing with a few sliders already yielded a few potentially interesting games. Having said that, with Steam I always feel there is a lack of a few filter options specifically to exclude things like early access games. Over time, I have found that most early access games either never get out of early access or are actually better, more enjoyable games when they do. Steam really seems to want to push early access games (or there are just so many of them that they dominate any listing).
LibraryThing gives me good book recommendations, and MovieLens gives me good movie recommendations.
Good grief, MovieLens is still around... I signed up in October of 1999. And it looks like the recommendations are still reasonably reliable, based on what I've watched since the last time I used it, maybe six or eight years ago.
I don't think the issue is that they don't benefit people in the short term. I sometimes benefit from them. Spotify recommendations usually do fit my tastes, and YouTube related videos have exposed me to a lot of interesting things. The problem isn't that in my eyes : it's the long-term, more discrete effects.
The existence of a recommendation algorithm on a platform, no matter its inner workings or its implementation, will inevitably push the media produced for that platform into one specific mold. The only thing you can actually do to make that less detrimental is to make sure that mold is a decent one for good content, but obviously, that's a subjective notion that will most likely be affected by the designer's opinions. (tbn: even without an algorithm, systemic factors will also sculpt the content being produced but at least those factors are directly observable)
And then there's the issue of privacy, since these algorithms rely on personal data collection, and the reliance on them encourages the practice to become more and more invasive.
But we're on Tildes and I'm rambling about crap that everyone here knows. The one question I still have about this stuff is... did anyone actually want recommendation algorithms in the first place? I'm not that old of a web denizen so perhaps I was just too young during the Web 1.0 -> 2.0 to know, but I don't remember ever wishing the Internet could magically know and give me things that could interest me, nor do I remember it being a big thing people wished for. I think we were all fairly content with what Google showed us, and we used to know sites tied to our interests that would lead us to what we might be looking for.
Shouldn't that be telling? It was never born from the end user's need, so it's not surprising it wouldn't be a net benefit for the average person. It's a tool for marketing and nothing else.
Discovery was (and is) a real problem for accessing new content; you don’t know what you don’t know.
A Web 1.0 tech that didn’t survive the Web 2.0 transition was StumbleUpon. At the time it was a breath of fresh air from the ensuing rise of link aggregator sites. It gave you seemingly endless random websites from parts of the web you would never have found on your own. After a while of clicking the “StumbleUpon” button (browser toolbars were also a thing that Web 2.0 killed…not lamenting about that one, though) you’d eventually exhaust the available sites and you’d cycle through ones you’ve seen before (i.e. “you’ve reached the end of the Internet”).
Eventually they introduced a Like/Upvote button which would help the cream rise to the top. In a way, this was likely a bare-bones recommendation algorithm (purely popularity based instead of using ML “personalization”). I can’t remember if they ever shared metadata like submission timestamp, number of upvotes, etc. in later years. I just remember it being a fun and charming part of the Internet that gave you access to the small web before it was solely personal blogs.
What I think made StumbleUpon’s charm so addictive was because it introduced unknown randomness to your internet browsing. Many times I don’t want to be recommended the same things I’ve already seen or what people like me have seen. Getting something to break me out of my comfort zone has value in itself and it’s something that is sorely lacking in the modern web. Everyone is stuck in echo chambers that encourages them to be xenophobic to outsiders and new ideas.
I don’t know if something like StumbleUpon could take off today. I checked online and its successor is an app called “Mix” (at least that’s who took over the domain name). After a cursory look, “Mix” is a video-meme doom-scroller. The memes themselves were ancient from what I could see. The network effects of Big Tech have entrenched such a stranglehold over information flow I don’t see anything that copies their formula butting its way in, either. Why would TikTok/Instagram/Reddit ever promote a competitor when it could tweak its algorithm to suppress and/or steal content instead?
Even the original StumbleUpon founder went on to found Uber, which makes me feel like his thoughts around a free and open web have likely changed since 2001.
Warning: potentially chaotic rambling thoughts incoming ;)
It was actually my dissatisfaction with StumbleUpon that lead me to reddit at some point. At first, it was everything you did describe, but it quickly turned into the sort of behavior we these days call "doomscrolling". Even made worse with the lack of interaction possible on content. Half of the things were blogs, with comment section that were at the time already terrible.
I also feel like that at some point the content became less novel and fun, but that could also just have been me using it too much.
Finding reddit at the time was a breath of fresh air. Reddit is still popularity driven, but at least it is/was much more fine-grained due to subreddits, and you actually had a community to engage with.
I am also not sure if StumbleUpon could take off these days, it seems people are less willing to move off their platform. You saw this on reddit as well, where content hosted off site often did less well compared to self posts and reddit hosted images and videos.
StumbleUpon got around that with the mentioned toolbar, even reddit had one called companion which I briefly maintained after reddit abandoned it.
Since browsers, even at that time, didn't have APIs anymore for toolbars, it worked by creating it with HTML on the page. Meaning it needs full access to all pages and even then it can't work perfectly, certainly with all the CSS and JavaScript trickery going on these days.
So on desktop it is difficult to make something that would work. On the mobile side of things, I feel like it could technically be made to work. But, it still is a bit of a hard sell, as people would likely want to see a listing of entries as well. At which point, you basically have created yet another reddit alternative.
Basically, all of the above is a long-winded way of me saying that StumbleUpon wasn't all that different from reddit. It easily could have pivoted to be a very similar platform. Which is likely why it would have trouble today, as any reddit alternative has trouble competing with it.
YouTube has a few things related to whatever it is you’re searching for followed by a ton of unrelated results followed by 1-2 related just to “keep you going” in my experience. Compare that with an Invidious instance and it’ll give you what you’ve searched for.
Instagram is another offender, although I can’t think of a Web 1.0 example. Make one mistake on the feed and it resets, forcing you to scroll again or search.
Closest thing to algorithmic recommendations was different radio station playlists IMO. what.cd in its heyday was great for recommendations via user collections. Letterboxd has a similar concept for movies without the actual content these days.
A lot of topics were alone on their own little “islands” usually filled by SMEs and enthusiasts plus the occasional person who registered for help.
Overall, I think things were just more organic and word of mouth vs everything having an agenda. But maybe that’s naive Web 1.0 me talking wearing rose colored glasses.
Youtube Music's has been pretty good, though it helps if you inject fresh blood from outside recomendations and periodically purge your history.
That said, YTM is an utter miscarriage of an Android app which is basically useless if you're not connected to the internet, even if you have dozens of GB of music downloaded/cached. It's so slow.
Remember when Netflix had contests for who could make the most useful recommendations? It's like we're in a different world.
For a while I got good value from the recommendations in Pocket. I would add a bunch of articles to read later, and what Pocket suggested to me from what other users have saved was often interesting reads as well. But that was years ago. However I don't think it is necessarily the Pocket algorithm that derailed as such, more like the open internet got even crappier and the well written articles ended behind paywalls.
For music, I am still finding new artists to listen to through Last.fm. A service that has been around for about 20 years now and thankfully hasn't changed that much.
A lot of them are good and bad. They don't work 100% of the time but they often get it right. A lot of what YouTube and Netflix suggest really work for me. If I don't care for something it's easy to skip, and there's often a "don't recommend again" button of some kind.
It's not that difficult for me to turn on "manual mode" and spend some time looking for stuff on the search bar of those services, as well as tools such AniDB, IMDB Advanced Title Search, AllMusic, Wikipedia, Reddit, Tildes, etc.
We used to have yearly books in print that were essentially like the phone book but for movies. They had a tiny sinopse and a rating for the main VHS releases for that year!
Because I am comfortable with options that circumvent many recommendation algorithms since decades before these systems even existed, I don't often complain about algorithms.
IMDB Advanced Title Search is so awesome, I'll be sad if it ever go away. I recommend it to everyone looking to easily build a highly peculiar list of movies to watch. It was recently revamped to look modern, but, fortunately, it kept all features which I love. IMDB is specially useful for old movies, let's say pre-2000 movies, because that is when users were more likely to be knowledgeable and things like "review bombing" were not really a thing. I doubt that newer websites will ever accumulate a lot of reviews for movies prior to the 1970s at this point.
So yeah, I don't really expect much from recommendation algorithms in the first place.
I can’t answer the actual question. Just share that I have the same problem with the almighty algorithms. I want a nice level of diversity in my options whereas they seem to want to just feed me basically the exact same thing.
Just as an aside Pandora used to be amazing. I like a lot of classics, and pandora was amazing for a short time at finding deep cuts i hadn’t heard that i really liked. I suspect it was licensing that did them in.
Yeah! My usage has dropped off a lot in the past ~5 years, so I was mostly speaking from memory, but before then it used to be really good. Back when Winamp was a thing, I think there was a plugin that tapped into the same Music Genome Project data to generate playlists from any given song, and I found that very good as well.
I worked there during the dot-com crash, before they changed their name and relaunched. The best employee perk was being able to listen to music in their collection while at work and explore based on the "genome".
Of course, they couldn't make full songs available to the public. They just sold the ranking algorithm to stores, with short clips.
Unfortunately they ran out of funding for a while. When I left I didn't bother exercising my stock options. It was rather surprising when they got new funding and came back again, streaming full songs by following the rules for radio stations.
Nowadays we have Spotify, which is a much larger collection, but their collection seemed really big at the time.
Pandora was the best for music discovery from the early to mid 2010's! It was dethroned by Spotify, which also used to have stellar recommendations a few years ago but is also becoming increasingly stale. It does seem like the natural trajectory for any media streaming service is to gradually steer users towards content that is cheap to license at the expense of quality personalized recommendations. That's one way to attain some short-term financial growth for the shareholders.
A rating/recommendation system really has to be independent from any media delivery pipeline in order for it to be effective in the long-term. If it can't be impartial than it's inherently flawed. Unfortunately we've lost most independent rating/recommendation platforms... most are dead or owned by Amazon at this point.
https://www.criticker.com is great for movie recs. Was just talking about it on another thread here.
Criticker was my first thought! I've been using it for over a decade and have found so many interesting, obscure films as a result. It has avoided enshittification after all of these years, unlike most of the other websites mentioned here.
Unfortunately it does still feel like the community is dying. Most recent titles don't have enough ratings to generate a score prediction. The recent UI overhaul was a huge misstep IMO. The old UI was very practical, even if it looked slightly dated. Now with the new UI, it seems like features break randomly or are significantly slower for no reason, and it's likely driving users away. It doesn't even look much better either- space isn't used efficiently and it takes more clicks to perform basic tasks. What a shame.
Lately, even the ones that were good adjacent have become off the rails weird
TasteDive is pretty good for recommending similar movies, tv, and music. As long as it isn't too niche.
Might be good to develop some good search habits like we did before the algorithm days. For example, if I want to find a similar movie, I'll search for various 'best X movies', and also enter my movie along side the search. That way I get around the listicles and fluff and it will be a list of movies that includes the movie I know I already like, so by relation I know the list will probably be similar. This is great for really niche things that won't appear in a generic 'top 10' type of list. You can also do this for products. Searching best X product and restrict your results to Reddit. Make sure to check multiple threads to avoid some company astroturfing their product. Find a general consensus and do a search similar to above, "Best products for what I want to do [Product mentioned on Reddit]". Another way of doing this is is typing X vs and letting Google/DuckDuckGo try to autocomplete the search. Then instead of searching type both products with the term 'list' or 'comparison'. Not completely foolproof but it helps avoid the marketing when all you want is recommendations.