I find a lot of sites feel like they're overtuning their recommendation engines, to the detriment of using the site. YouTube is particularly bad for this - given the years of history and somewhat regular viewing of the site, I feel like it should have a relatively good idea of what I'm interested in. Instead, the YouTube homepage seems myopically focused on the last 5-10 videos I watched.
The problem is the economics of the Internet today. Most sites are ad-funded. They need to maximize pageviews and time on site. The recommender is a huge part of accomplishing that, and it simply won't be tuned to any metric other than maximizing revenue. The solution would be to reskin popular sites with recommenders that had other objectives, such as retrospective satisfaction with time spent. Unfortunately sites know that if they give up control of the UI through an API that is open, they will lose money when people do things like this.
I really think we're entering a crisis here where sites and apps that are engineered to maximize corporate metrics are leading people down horrible paths of addiction, and psychic stress as they spend their mental energy resisting temptations constantly thrown at them.
Thought just occurred: would I be willing to pay a subscription to have the recommender tuned to remove revenue maximisation and site-addiction maximisation? Would anyone?
I've made plenty of "remove ads" in-app purchases on my phone. This isn't too different. And it might actually result in a truly useful experience.
There may be public relations problem - users these days understand that ads pay the bills and don't see them as a moral compromise but try to explain to your users that they can pay premium to get relevant recommendations instead of "spam".
I've said the same thing many many times, I can't remember the numbers but the monetization of facebook data per person (If I remember correctly from something I read a year back, I'll try to look and add it to this comment) is less than 20$!
I would be very very willing to pay for Facebook to simply not be tracked, especially at such a reasonable price. I know that won't happen, but I'd really love such a thing.
It's only so low, because those are averages. The people willing to pay the twenty bucks might be exactly the ones that drive up the average, so they don't want to lose them.
That incentivizes using adblockers though, and given that ads are pretty much YouTube's entire business model, that would be nice for you but terrible for the business.
However, they do charge money per month for YouTube Red which removes ads, and a better recommendation system for Red subscribers might encourage people to buy it.
I'm not sure if it is because of weird tastes or something else. But for sites like Steam, YT, Netflix I always wish for more tuning parameters because their recommendations are all horrible.
On Netflix, for example, I get recommended stuff that's similar to other things I didn't like.
Amazon is often the best of them for books because their engine seems to only really value the last few things I looked at / bought and so just recommends really similar stuff. Of course, that fails hard whenever I look at something I'm not interested in for whatever reason.
On Steam I wish they had some elaborate filtering system, there are so many games, I need something like NOT(Adventure+Puzzle) BOOST(RPG), I need to combine tags, not just filter them out by themselves.
> for sites like Steam, YT, Netflix I always wish for more tuning parameters because their recommendations are all horrible.
The problem with those recommendation engines is that they're not optimized to serve your needs. They're optimized to serve the goals of their respective companies. And the problem comes when your needs are a little incompatible with the needs of the company.
Consider Netflix for example. Their recommendations don't seem to care much about what you actually enjoy the most. They pay different amounts of money to let you watch different content. So your goal of watching the thing you would enjoy the most is different from whatever the hell their goals are-- probably to get you to watch just enough Netflix to make you not want to cancel your subscription, but not run up their bills on bandwidth and licensing fees.
I'm absolutely confident Netflix could make amazing recommendations-- and probably already has them internally. But it's not in their best interest to give recommendations that are in your best interest.
To the extent that I'm right about this it could be a market opportunity to make an honest and useful recommendation service.
Netflix has a tiny catalog, so investing in making recommendations from that tiny catalog is mostly wasted. Half of Netflix's business now is House of Cards, so their entire algorithm consists of recommending House of Cards to people who haven't watched it yet, and random content to the others.
> They're optimized to serve the goals of their respective companies. And the problem comes when your needs are a little incompatible with the needs of the company.
Ultimate they are compatible though. And yes I mean in the econ 101 way. This tension should be at least partially resolved via pricing scheme innovation.
Steam would earn more from me if I did not closed it frustrated every time I wanted to buy game. I literally did not bought game because there was no way to filter out stuff I don't like (well defined category in this case).
It's worth noting that satisfaction is also a function of perception. Did you enjoy the movie? Yeah, probably. But then you think about all the possibilities, a small number of which would naturally be better, and an acceptable film fades towards unacceptable.
Regardless of how good the recommendation engine might be, it's still always subject to human perception. That can be unpredictable and fickle, on a good day :)
I once clicked on a 'russian bride' google ad for a laugh, quite a while back. Google still puts lots of russian and filipino bride ads into my 'bubble' to this day...
Not the same person, but I have three. I have a roomba for everyday sweeping, a Bissell for actually cleaning the carpets not just sweeping them, and a Black and Decker handheld for the stairs, the car, and other places it's hard to take a full vacuum or a robot.
My Amazon homepage is full of stuff I already own. Even worse is their internet-wide remarketing makes it so I see stuff I own on many of the sites I visit.
There was however one moment where Amazon got it right: they recommended to me a book that only a week earlier I had purchased on a whim from an independent bookstore with cash. Creepy good.
> There was however one moment where Amazon got it right: they recommended to me a book that only a week earlier I had purchased on a whim from an independent bookstore with cash. Creepy good.
That is creepy. Any way it might've been more than just a coincidence?
It's possible, but it was a book in a relatively unusual niche that I have a hobbyist interest in. It was a book about the archaeological and linguistic study of proto-Indo-European people (PIE) and their origins and spread by means of horses and the, at the time novel, wheels. I had previously purchased a book or two on linguistics from Amazon so perhaps there is only a small set of popular books in that category for their recommendation algorithm to pick from.
I told this story in the recent Amazon recommendations discussion, but I bought a pair of speaker stands from them and before it was even delivered they sent me an email saying "HEY BASED ON YOUR BROWSING HISTORY MAYBE YOU'D LIKE THESE OTHER 10 SPEAKER STANDS!"
wait, wouldn't this cause buyer's remorse? like, if I just finally decided to buy some product ... being bombarded with all the other options suddenly makes me rethink my purchase decision, "oh but what if this other one is more gooder!?!?!"
I think the argument is cognitive dissonance and confirmation bias generally function to make you more happy with the one you bought already when presented with other alternatives.
I thought that cognitive dissonance was the buyer's remorse. I've experienced that a few times in my life. Now I've got no qualms about immediately taking stuff back.
A while ago I took an Ayurveda class (Indian medicine) and part of it was doing an oil enema. There are devices for this that are also used by a certain population for sexual pleasures. Bought one on Amazon and for months I couldn't show my Amazon page to other people because it was filled with sex toys....
My favorite sleeping mask is made by Joy Division, a blindfold sold as a sex toy. When I replace it every 1-2 years (or earlier in case I was on a flight…) my Amazon page is full of sex toys ;)
For books this works pretty well for me as I tend to read genres in bursts. If I just read a scifi books, I'll read a bunch more before going back to fantasy ;)
The equivalent for books would be "you might like this paperback copy or a large print version of the book you just purchased along with 3 other editions!"
It's probably correct by expected value. You're just not the multiple wallet type. I'm sure there are people who'd want five wallets, eight different shoe styles, multiple cuts of pants...
Your indifference is subsumed in the sheer size of others' consumption.
You're right that there are categories of products that this absolutely works. When I buy clothes, books, media... even tech, then I probably will buy something related. But there are some "one-off" purchases that Amazon doesn't need to hound be about.
I have multiple wallets. I have one I take on business trips that contains my work ID, American Express, business cards, etc. All stuff I normally don't carry with me. I have an everyday wallet with cash and my bank card and stuff I need everyday. And if I'm dressing up, I don't want to ruin the lines of my suit so I have a far slimmer wallet that just has my ID and space for one credit card.
I mean I'm not accessorizing with the latest seasonal fashions, but there are utility reasons to change the kind of wallet you're carrying for different purposes.
Netflix, I think, has killed their own recommendation algorithm when they removed stars and made it boolean. I don't know if those buttons even do anything anymore because I think they're just matching based off demographic and who they're trying to market to now. It's recommending shows that I never would watch in a million years and giving them high matches despite me disliking most similar shows just because I'm a 20-something male.
I think Netflix probably did it because people are inherently bad at being objective. For the average person, 2 stars for a movie and 4 for another isn't based on anything measurable, even they couldn't explain. I'm shocked at some of the amazon product reviews, most of which are 5 star reviews even if the product is absolutely terrible. Movies are different than products, but it's the same people doing the reviewing. Remember, the average user is not a thinking analytical HN user. Average people are much better at bool choices.
I know I'm in the minority here, but I am a big fan of the new system. I would torture myself trying to decide between, e.g., 3 or 4 stars for movie. And then go back and re-rate other movies that I realized I liked more but rated lower than the just-rated movie.
Their % match numbers are fairly accurate, but I have had to go into the watch history and delete the occasional movie watched and finished that we actually hated. No number of 1-stars (or thumbs downs) would eradicate its effects on the recommendations.
5 - Absolutely loved it, will buy a disc
4 - Good, but won't buy a disc
3 - Movie was okay
2 - Not a good movie
1 - Stopped watching 20 minutes into it
My problem with binary choice is that 1 == 2 and 3 == 4 == 5, whilst 1 and 5 were very special for me. :(
Plus the scale bias differing vastly between people and cultures makes the data a mess. Like say or me a 5 means 100% perfect, Why discreet choice stuff is all the rage in the market research world. (unless that's changed in last few years)
Asking people "which of these 3 things you like best" vs. "rate these 3 things 1-5" will usually give you much more useful data, plus be easier for respondents.
Popular recommendation algorithm like collaborative filtering by matrix factorization takes into account the accounts for user and item biases (the simplest method is to normalize the ratings of a particular user by the average of ratings of that user).
Couldn't you control for that by weighting people's ratings by the range in which they provide them? Like weighting a 5-star review a bit more from someone who averages 3's than someone whose ratings average 4's? Far from perfect sure but I bet it could save a lot of results from needing to be thrown out.
With stars you can cross compare with others to see if they have the same score. With simple thumbs up recommendations you cannot compare the ratings as the score is whether it appears to you or not.
I have to wonder if Netflix did this because a lot of their original or exclusive content seems to debut to mediocre star ratings. When the new system says "x% match" I assume that value is derived more from genre match or search relevance than whether I'll actually like it or not.
"In addition to the new rating system, Netflix has new match percentages (up to 100%) to more accurately predict how much users will like something.
...
The new rating system received 200% more ratings in A/B tests, according to Netflix VP of Product Todd Yellin.
When it comes to rating movies and shows, stars reflect the preferences that people want to have, rather than how people actually behave. Todd gave the example of users giving 5 stars to a documentary but just 3 stars to an Adam Sandler movie that they watched over and over again. “What you do versus what you say you like are different things,” said Todd."
If Netflix was trying to ensure that what I was mostly likely to watch next had the highest star rating, no wonder they gave up on it. Our opinion of the quality of something is not a good predictor of our likelihood of watching it.
Their users are surely not confused about that. So why does Netflix want to present a prediction as a rating? Is it to flatter their users by telling them that the thing that feels instantly gratifying right now is actually an amazing movie? "Hey, great choice. Billy Madison is a five-star movie. What? Why would you feel bad about not watching Raging Bull instead? It's a two-and-a-half star movie at best."
In other words, Netflix, like Facebook, like Doritos, is engineering itself for maximum addictiveness without regard for honesty. It will shelter you from even what you know and reassure you that whatever triggers a pleasure response in your brain is the best. Relax and enjoy it.
The truth is that we consume easy things a lot more often than we challenging things. It would be exhausting otherwise. But the best things are often the most challenging. We know that, we know that easy movies are just a way to kill time, but Netflix wants to do us the service of helping us forget it, because then we might be 1% more at ease when we watch Netflix and 1% less likely to switch to another service.
>When it comes to rating movies and shows, stars reflect the preferences that people want to have, rather than how people actually behave. Todd gave the example of users giving 5 stars to a documentary but just 3 stars to an Adam Sandler movie that they watched over and over again. “What you do versus what you say you like are different things,” said Todd."
There's a reasonable objection to this behavioral definition of "like", which is that it doesn't actually make people's lives better, it just fills them with more compulsive behavior. It's not necessarily "irrational" to wish you were more patient, or to want to ask Netflix to show you useful things rather than useless fluff. That you occasionally betray your stated goals does not mean you should be denied the right of self-definition.
In other words, what people say they like is more important, to me at least. See eg "Thinking Fast and Slow" by Kahneman.
I can't agree more - I'm a huge proponent for star rating systems. I get that they are perhaps more complicated than a Boolean value, but they help me out personally.
I miss the days of "tap tap scroll four clicks" on the iPod to help me rate new music, specifically.
Thumbs up/down might not be the best for training a recommendation engine, but as someone who just switched his product's rating system from 1-5 to simple up/down, let me tell you: people have no idea how to use a star rating system. I would get people raving but leave a 1 star review, some people would leave a 5 star review and say bad things, some would leave a 1 star review but seem pretty neutral in their review.
All the technology hype aside, it often feels as if these feed prediction algorithms are akin to weather forecasting. That is, there not forecasting per se, but taking educated guesses based on some set of knowns. The problem is correlation is not cause.
Part of that trend is because recommendation algorithms aren't all that interested in making recommendations anymore. They're interested in getting sales or views. To do this, they use dumber algorithms that are easier to understand ("people who bought this also bought...") and over-value recency. They're not helping you find new content, they're trying to prolong the time you're on youtube.
I have been paid to write an algorithm like this. It matched you to relevant content, but then a manually configured weighting table would skew results to more popular sources. It went as far as taking into account the number of leads we had sent each partner that month as a percentage of their quota
Youtube is definitely over-weighing the last few videos. Then again, it's likely optimized for a different market segment. Especially with how many children are given tablets. Like, brief intense fascination with many subjects in general is genuinely hard to optimize for, especially when a huge portion of your market has long-lasting intense fascination with a few subjects.
This is often the behavior I want. I frequently queue up a song on YouTube and then let it keep auto-playing, if I hit one I don't care for I usually select off the sidebar of recommendations.
This only works because YouTube stays focused on what you're doing without trying to build a bigger picture of who you are. It would be nice if it could somehow do both, but I don't have a clear picture of how that UI would function.
YouTube's home page system is quite hackable in a way, but you need to use the search or other recommended videos to change your viewing pattern.
In other words it doesn't consider you as someone with a long history. You can change your profile from a conservative to a liberal in a few hours watching videos. Whether it's possible to have a balanced amount of crazy ( not the same as a centrist) is something I'm currently working on, it requires effort!
YouTube would rather you fall into a trance of continuing your recent viewing habits rather than providing a personalized library. It just makes better business sense.
Because they aren't trying to help you enjoy videos over the long-term, they're trying to keep you from going over to Facebook where their addictive algorithm will take over and reduce your likelihood of returning to YouTube!
It's a classic race-to-the-bottom. If YouTube lets Facebook keep you addicted, you won't get to the long-term.
Yes, but that has an unspoken assumption that recent habits are more additive than longer term ones. I find that suspect; after all, you've already shown that you'll go back again and again to these kinds of videos, whereas you might spend a couple of hours watching a bunch of videos of a kind that you'll never be interested in again.
Personally, I find YouTube terrible at keeping me engaged. Even when I want to keep watching videos, their sidebar fails to show me anything I want to click on at least 1/3 of the times.
Yes, but that has an unspoken assumption that recent habits are more additive than longer term ones
This is too fine a point. A user will "select" a default choice by doing nothing, and it's in YT's interests to provide default choices that reward the company. This has been researched for decades and is supported by, among other things: https://en.wikipedia.org/wiki/Status_quo_bias
Because if e.g. you watch a Tool video, the ad rates for you watching a Korn video after that are higher than if you watch a Zeni Geva one. Alternatively, Kesha > Lady Gaga vs. Kesha > Joni James. Unscientific, but it comports with conventional wisdom with regard to fiduciary duties to stockholders.
Because if e.g. you watch a Tool video, the ad rates for you watching a Korn video after that are higher than if you watch a Zeni Geva one
But why? If I've watched hundreds of Zeni Geva videos over the years, but it's my fifth Tool video, why would I be more likely to click on a ad on a Korn video?
It's not that they think you'll be more likely, it's that they rather you did, so the suggestion emerges. Recommendation engines are a funnel for revenue.
As largely a geeky population, perhaps we do need to accept that our interests are more varied and often actually more intense than those of the average person. Being mildly amused by every fart joke movie available for some people is more rewarding than watching an emotionally challenging drama, a thoughtful comedy, and eight documentaries about four topics.
It is not all that bad. If I listened to three metal songs last few videos, I am likely to want some more metal in the next one. I am unlikely to want jazz or comedy sketch or kiddy cartoon (despite these being seen from my account last few days).
> Instead, the YouTube homepage seems myopically focused on the last 5-10 videos I watched.
Noticed the same thing happening for the past year or two.
I used to be able to go on YouTube and find a variety of interesting videos I'd never seen before. It seemed like there was a good balance of categories in the recommendations.
Now, I watch one boating video and suddenly my recommendations are 100% boat related with some random clickbait/viral garbage sprinkled in for good measure.
YouTube is weird. I made the mistake of watching a flat earth video last year in order to get a handle on that growing weirdo fad. Well it turns out that flat earthers are really avid content consumers (presumably because of the enormous cognitive reinforcement requirements to maintain such a belief) and YouTube really wanted to help me with that. It took a couple of months for it to stop offering me a portal to a better, flatter world every time I refreshed a page.
Once a video is removed from your history, you'll no longer get recommendations based on it.
If you aren't logged in, you can't view or edit your viewing history, but you can at least clear it. (Which will return you to YouTube's terrible default viewer profile... sigh)
I similarly had that happen with some GamerGate stuff. Watched a video just to see if I could wrap my head around it, went "Nope!" and went on with my life, and the next thing I knew YouTube was desperately trying to send me down the alt-right rabbit hole.
It's a hard problem. If they kept recommending old topics, you'd be like "hey, I quit watching pink zebra videos 5 years ago, stop recommending that stuff to me".
My supposition is that this is a response to changing tastes over time. However, I think 10 videos or less it probably an overreaction to this problem.
For example, let's say they have a user who watches 100 videos a week, for ease of math. 50 of those videos are in "core" areas of interest - these do not change over time. An additional 35 are in areas of secondary interest: topics which have piqued the viewer's curiosity, but not deeply interested them. We can expect these topics to change every [1,4] weeks. The remaining 15 are referrals or clickbait from other websites.
How can YouTube differentiate between these three classes of videos? The first class will be heavily represented in their subscriptions. Presumably, the viewer will prefer their recommendations to ignore the third class (clickbait). The second class is the hardest, as the user may want these videos surfaced, but then want them to decay over time as their interests change.
I think this is the problem that they are attempting to solve, with varying degrees of success.
It's not just over tuning, but finding me things I might have missed. Not necessarily popular things, but thinks that appeal to me.
For example, I get emails from Pocket "You saved a popular article..." And I think "Who cares?" Don't tell me what I know, tell me what I missed. YT is similar. It recommends things I've watched. I'm looking for new and interesting and I'm getting yesterday's news? That doesn't excite me.
I hate this. Several years of nothing but a single genre of music in my history but watch one video about something different and get nothing but recommendations based on that. I have to clean up my watch history daily.
I've moved pretty much all my music listening to YouTube due to simple convenience of easily being able to generate whole genre playlists based on one song.
But because the playlists are dynamic, YouTube keeps on shoving songs from other playlists (and genres) into each other, trying to generate a "perfect playlist" and in the process making all the playlists sound very similar with no more genre distinctions except for the first couple of songs.
yeah, but i get the impression history is retained and utilized. If i watch 2-3 how-to videos i start getting how-to videos that are similar to my history, not just random how-to. then if i switch to skateboarding, i get skateboarding videos similar to my history.
This seems like a good algorithm to me, as when i'm watching skateboard videos with my friends, I don't want "how to caulk tile joints" to show up
Well, it sometimes makes sense. I, and many others, use YouTube as a music streaming service. Most people listen to most songs they like more than once. I dare say yt has gotten a good amount of ad revenue out of me listening to Darren Emerson's dub extravaganza mix of black sky by Shakespeare's sister, which is always in the recommendation panel for some reason...
I concur that.. these days I have to clear my youtube history every few weeks to prevent it from spamming my page (and my fucking TV!) with suggestions related to random videos I watched over a short period of time. The videos I liked over the last 7-8 years seem to carry less weight. I wish there was a way to tune these behavior.
If I had to guess, I would say YouTube is attempting to get you to watch new types of videos you've never watched before. By focusing on recent videos you've watched, they can try to convert you from a 1-off to a regular viewer of a particular genre.
i think this is the right approach, you're not after all, the person you were 5 years ago, or even a week. most interests change with time, I lose interest in a topic very fast, and
gain interest in another topic just as quickly. No good recommender system should be based off a reading of my "personality", whatever that may be - the most stable aspects of my personality, even if they can be divined from my viewing history, say little as to what I would be interested in watching next.
I think that's a fair point, and I wouldn't expect a video I watched 10 years ago to factor in very heavily on what I'm seeing today. But at least in my experience, it doesn't appear that the engine takes anything that happened more than a few days ago into account.
As an example: I watched a few episodes of Penn and Teller: Fool Us yesterday. I hadn't really watched it before, and while I like Penn and Teller in general, I don't remember watching them all that much on YouTube prior to yesterday (I'm sure at some point I had watched a video or two, but not more so than anything else I stumbled on.)
Today, 12 of the top 30 videos on my YouTube home page are specifically Penn and Teller: Fool Us. Not magic in general, not Penn and Teller in general, but specifically that show. That seems like the very recent past is way overrepresented.
This is because they're not trying to give you stuff you'd generally like given your entire history. The recommendation feature powers the part of YouTube that auto-plays the upcoming video.
So e.g. if I go and view Russian dashcam videos they're going to automatically play more of them, even though I've shown no prior interest in that topic.
Having two systems for recommendations would introduce a lot of UI complexity, so I can see why they didn't go for that, and why the recommendations are consequently tuned for people who are actively watching videos on some topic right now.
Having two systems for recommendations would introduce a lot of UI complexity,
Do you mean, in terms of implementation or for the user? Because as the later, I think the two are already conceptually different (homepage vs next video), so I don't see how just feeding it different videos would make it more complex.
For the user. What you're describing is still going to cause complex UX.
So let's say I'm interested in exactly two things. Russian dash cam videos and videos of trains without narration.
I go to the homepage and click on one type of video, what should the next video be? A random video from the two categories? One or the other? Should it show me a banner at the top indicating what recommendation mode I'm in (historical preference or "similar to current video").
Now I go and do the same on my Android YouTube app which has no real UI equivalent of a homepage, what happens then?
This is a lot of UX complexity for a feature few probably care about.
Like I said, I think the homepage is already quite different from the next video in the user's perspective, so there's no need to show banners or anything like that. Homepage = full profile analysis, next video = similar to current.
By the way, the Android app does absolutely have an UI equivalent of a homepage, it's even called "Home".
I mean once the user selects a video you need to continue showing them context, because at this point they may not remember what mode they're in to begin with.
Making your app behave differently because you navigated to the current state via different menus is very bad UX design. That's all I'm saying. Your suggestion would entail either UX complexity or such implicitly different behavior for YouTube.
Yes Android has a "Home" button. But what I meant by no real UX equivalent is that when you open youtube on the web you'll open a new tab and go to youtube.com.
When you do so an Android you've likely just dismissed the app in the past, and opening it again will bring you back to the last video you were viewing. You don't go through the homepage by default.
Thus it's more of just another menu item on Android, not something that's equivalent to / on a website.
I imagine it would be possible (though difficult) to autoplay videos which are related to each other and gradually converge to something that would interest the viewer.
For example, Russian dash cams to Russia at night to the Russian sleep experiment creepypasta, to horror games to video games in general, if that is what the user tends to watch.
I know this has graph theory written all over it and the shortest-distance problem has wreaked havoc for centuries, but I think with enough resources Google/YouTube could find a good compromise in this situation.
