9 ms·

I work in this industry and actually we’ve found from online experimentation customers generally don’t want to tell you they dislike something. The expectation

I work in this industry and actually we’ve found from online experimentation customers generally don’t want to tell you they dislike something. The expectation is customers get what they like (relevant) and sometimes they want to express stronger affinity for specific things. The converse hasn’t shown to be true. Generally customers don’t have a clear picture of how their explicit signals impact recommendations. Though some customers have shown desire to want to manage these signals. Edit: one other consideration is disliking can usually be inferred from other activity. Skipping a song after less then 20 seconds for example can train a model as well as incorporating an explicit dislike button.

marginalia_nu 5y ago
I had to keep deleting my Spotify account because my recommendations kept getting contaminated with music genres I didn't like that were impossible to get rid of. Listening to the wrong song two times could irrevocably ruin my recommendations. I would have very much liked a dislike button, or at least a "forget everything you think you know about me"-button. Now it ended up with me getting rid of Spotify, because of this and other frustrations.
- notyourwork 5y ago
  Without specifics I cannot guess as to your specific case but in general customer's just do not use dislike buttons. So exposing them isn't a value add and usually a UI clutter that most customer's have no use for and won't use. In general, you had to be doing something to be getting recommendations so there is some action that when most customer's do demonstrates a high probability that you will like some genre and it just so happens you don't.
  - chillingeffect 5y ago
    What industry just out of curiosity? I see dislikes used judiciously in youtube and reddit fwiw
    - notyourwork 5y ago
      Music but I don't work for Spotify currently.
  - marginalia_nu 5y ago
    > in general customer's just do not use dislike buttons. So exposing them isn't a value add and usually a UI clutter that most customer's have no use for and won't use. A feature does not need to be used often to be useful. Take a fire extinguisher for example. Going by this logic, since most people will never even touch a fire extinguisher and most of them are never actually used, we should get rid of them because they're a cause of clutter.
  - feanaro 5y ago
    I must say your dismissal of this is rather annoying. Multiple users are telling you they are facing a real -- and completely obvious and predictable -- problem due to the feature not being available. In a comment section regarding research showing it would be an improvement to recommendations. Increased UI clutter due to a single extra icon is not a good nor sufficient argument for not including it.
    - notyourwork 5y ago
      My dismissal is based on a many-million digit customer base and online real world data. Sorry if a few hackernews users feel they want a dislike button, most users don't and if you give to them most users don't use it. I don't mind disagreement but dismissing real world data from this industry is sort of silly to counter with 'because I want it you're wrong'.
      - feanaro 5y ago
        I'm not even saying you're wrong, I'm saying your argument for not including it for (at the very least) a sizeable minority is weak and somewhat silly.
        notyourwork 5y ago
        What is a sizable minority aside from an oxymoron?
        hulahoof 5y ago
        A sizeable minority is something that is still visible on a pie chart.
        feanaro 5y ago
        A minority is anything that is not a majority (this is the definition I'm using). So a < 50% group can hopefully still be viewed as non-negligible, even though it is not the majority.
      - 5y ago
  - breakfastduck 5y ago
    I absolutely love it when people working on a product tell their customers that they don't need a feature they're all asking for and that would clearly improve the experience, just because not everyone uses it all the time.
    - simion314 5y ago
      > absolutely love it when people working on a product tell their customers that they don't need a feature they're all asking for and that would clearly improve the experience, just because not everyone uses it all the time. We need a name for this, personally I call it GNOME mentality but probably is bad name since most developers are not familiar with GNOME (though the file picker issue get on top HN a lot )
    - notyourwork 5y ago
      > they're all asking for All is the flaw in your comment. One person does not imply all and data speaks for itself.
      - marginalia_nu 5y ago
        The data does not speak at all. You are drawing inferences based on the data, inferences of a type that are pretty easy to poke holes in by looking at the pretty large class of things rarely used but situationaly useful (often life saving). This type of bizarre logic makes me feel like I have to habitually use features even when they are not useful in order to protect them from UX designers in case I'm subject to some misguided A/B test.
        breakfastduck 5y ago
        UX designers not sufficiently held in check by strong product teams are lethal.
        notyourwork 5y ago
        > The data does not speak at all. What data are you referencing? I'm drawing conclusions based on data I have but I don't think you work for the same company as me so I'm not sure how you can speak to confidently about how wrong I am.
  - toofy 5y ago
    Each of the comments that mirror yours illustrate a major problem we have in the tech sphere—assuming _we_ know better what the end user needs therefor we don’t listen to the actual user. In our defense, we’re very disconnected from the end users. I’ve done some work with a couple disaster organizations in the past and we continually get massive praise and thanks from the victims of these disasters. The actual victims have the same praise over and over again: Unlike the red cross, salvation army, and FEMA we actually ask them what they (the victims) need. We don’t assume they need water and and a tarp, we assume they know whether or not they already have those things. Often the victims need diapers, help with flood mold remidiation, a chainsaw to cut limbs from their street so they can get their cars out so they can go help the neighbors a few streets over etc…. They know what is prohibiting them from getting out to help others much better than we. When people are telling us what they need, we should listen to them and we should actively try to avoid our tendencies towards “I know better what you need.” I have a couple friends who work on recommendation algorithms full time and in their defense, they’re fully aware how terrible the recommendations are—I still get my music recommendations from friends, from the local music store employees, and friends who work in music venues—their recommendations are infinitely better.
- the_only_law 5y ago
  > Listening to the wrong song two times could irrevocably ruin my recommendations. Similarly any app or website that pulls “top artists” from Spotify tends to include stuff I listens to a bit several years ago, but never stuff I’ve listened to nonstop for months.
- 0134340 5y ago
  Yes that's very frustrating. I'm scared to even listen to new music on it because I'd try to use it for discovery but it would mistake what I listen to for what I like. Besides that it hardly recommends me any new music anyway. I now just use youtube for discovery.
  - marginalia_nu 5y ago
    They should add an incognito mode or something. You could even do this maliciously. Send someone a Spotify link to a Steely Dan song and a Skrillex song, and now yacht rock and dubstep will follow them around until they delete their account (I'm assuming the intersection between those fan-bases is vanishingly small).
    - Oddskar 5y ago
      That already exists. Profile > Private Session
      - marginalia_nu 5y ago
        Doesn't that just hide your activity from other people?
        Oddskar 5y ago
        From their support docs: Note: Anything you listen to in a Private Session may not influence your music recommendations, e.g. Discover Weekly.
    - vanilla_nut 5y ago
      High school me would have been very offended to be labeled “vanishingly small.” But you’re not wrong. By the way, I believe Spotify does have a private mode, which at least blocks your plays from showing up in the social pane and on your profile. But I’m not sure if it cuts them out of your recommendations. Please, dear God, Spotify: just because I listened to the Hamilton soundtrack once does not mean that I want to hear it every week in Discover Weekly and multiple times per Duo playlist, because Lin-Manuel Miranda is apparently a genre of himself.
- 1_player 5y ago
  I keep hearing from people how good is Spotify "Discovery" playlist. I've been a paid customer for 5+ years and their recommendations have a hit rate of 2%. I listen to a lot of different genres, artists, mostly dictated by my mood. Today it was some easy Dire Straits song, yesterday QOTSA, the other day some psytrance. So Spotify recommends all genres in the same playlist, creating an awful cacophony of songs that don't flow into one another. Spotify recommendation system is very bad in cases like mine apparently. It's so bad I keep listening to the same old songs instead of discovering something new, which is why I'm no longer a customer until they improve their AI.
  - breakfastduck 5y ago
    I listen to a LOT of different genres of music and their Discover reccomendations (not the playlists, the actual you like X artist you might like X) are absolutely brilliant. Got to be approx 50-60% hit rate. It seems to differ wildly between people - I've heard plenty of people making the same case as you, but many that have the same as me. It never appears to be 'its ok', its either 'brilliant' or 'terrible'. I've been using Spotify for 10+ years at this point, so maybe that makes a difference?
  - newsbinator 5y ago
    Same here. I like Christopher Tin (e.g. Baba Yetu) and therefore for 6 years I've had no way to tell Spotify I don't like 8-bit video game music. No amount of skipping or "I don't like this artist" seems to convince Spotify to stop offering me variations of 8-bit Tetris themes or MULTIPLE 8-bit Cowboy Bebop Tank! theme variations per every single Discover Weekly. I'd be thrilled to even get a 2% hit rate on music I like. Seriously Spotify: two separate "Tank!" versions per Discover Weekly is too many.
    - kortex 5y ago
      Man, the spotify algo makes no sense. The other week I was in a really specific mood, starting with Sonic 1/2 OST and moving into 8bit/chiptune/keygen. I probably listened to every rendition of Chemical Zone out there. Rest of the week, much the same. Nothing in Discover Weekly.
user-the-name 5y ago
> one other consideration is disliking can usually be inferred from other activity. Skipping a song after less then 20 seconds for example can train a model as well as incorporating an explicit dislike button. This is nonsense. The fact that I skip a song does not mean I don't like it in general, it means I don't want to hear it right now. Music taste changes with mood and situation. If you treat a skip as "I don't ever want to hear this again" you will collect absolute garbage data, and if I know you are doing this, it will make your app unusable because I can't do normal everyday actions without ruining my recommendations.
- notyourwork 5y ago
  It depends right. If you skip a song everytime its played in a sequence is a stronger indication of dislike compared to skip it once but listening to it a few times. The point is there is more to context and recommenders than a adding a dislike button and having a magically awesome recommender.
  - oriolid 5y ago
    Have you read about illusion of control (search for elevator door button for one well known example)? If we take the idea that recommendations are really hard and dislike button doesn't help at face value, wouldn't it still improve the user experience to give them the illusion that they can directly affect the recommendations? I'm not really convinced about recommendations being hard, though. Pandora could do them a long time ago, Netflix had good recommendations at one point but not since they dropped the star ratings and at some point Amazon wasn't that bad either.
rednalexa 5y ago
Is this within a global context or a local one? Liking a song in Spotify is global - Spotify then adds it to a list of liked songs. But from my experience with Pandora, radio stations were localized. Liking a song did add it to a list somewhere, yes, but disliking a song just meant that a particular radio station on Pandora wouldn't have a song in that particular style, and other stations did not seem to be affected by this. In other words, what you want to listen to, for me, seemed to be very contextual. And since I felt my dislikes were captured in a context (radio station), I didn't have any qualms disliking frequently using Pandora. And my experience with Pandora for song discovery has always been much better.
ozzythecat 5y ago
> Skipping a song after less then 20 seconds for example can train a model as well as incorporating an explicit dislike button. How do you know I dislike the song? Instead, maybe I just heard it (or a similar song) on your platform or elsewhere outside of your app and don’t want to listen to it again right now? I guess you could try to build a more complex model that determines what songs or music I like at what times of the day or what day of the week, or based on what I just listened to. IMO this is rampant in FAANG companies. We create complex solutions to problems that didn’t really exist and pat ourselves on the back when our experiment show positive results. The experiment might show you some metric (total time in app, play time at whatever percentile, etc) increased. Great! Now we launch the model in production for all customers and pat ourselves on the back. Meanwhile, the actual experience became shittier. A few promotions come out of this, and we continue living in the bubble. Alright, end of rant.
- notyourwork 5y ago
  The paper is making an obvious claim. If I have data showing positive and negative affinity, that will perform better than a model which just has positive day. I don't find this very surprising but my point is in practice exposing dislike buttons to customers are a CX that is hardly ever used. So the point that a model will perform better with data that customer's rarely offer up to your platform makes their claim moot in my opinion.
  - corin_ 5y ago
    I downvoted your comment to demonstrate that I disagree with you. I've heard some good arguments against allowing users to show dislikes/downvotes as well as likes, but I've not heard your argument before and it doesn't strike me as accurate. Less used than the positive direction, sure, but still significantly used when available.
    - notyourwork 5y ago
      Netflix, Youtube and reddit aren't the same medium as short lived music streaming is. Streaming music is a show lived precision game of guessing what's relevant. There is a lot of it depends scenarios, are we showing recommendations on a landing page, are we sequencing songs in a station or are we ordering results on a search page. The CX context is a big part of how you can use signals to benefit the customer. My argument isn't an argument so much as a from real world data in this industry at a major industry competitor that customer's don't use the dislike button enough to justify it taking over your CX. Using co-occurrences of playback across a customer segment and combining that with likes or other signals is good enough.
Groxx 5y ago
Skipping doesn't imply dislike, though it does correlate (because the reverse is generally true, people skip songs they dislike). But I routinely see skip mentioned as non-downvote sources of dislike information. Easily half of the times I skip it's because I don't want it right now. Some of my favorite songs reproduce absolutely horribly in a car on a highway for example, because the surrounding noise utterly destroys subtleties. So maybe you add "but it's not a downvote if it's hearted or otherwise positively ranked"... but many other times I'm just not in the mood for [a song I had never encountered until now].
- anticensor 5y ago
  Correlation does not imply causation, and neither does causation imply correlation.
ics 5y ago
A problem I have with the inferred dislike is that when I skip a song it's only something I dislike half the time. The other half I'm simply not in the mood to hear it, maybe for another 10 minutes or another 10 months. The result has been that I've become more willing to listen to certain genres on another platform (say Youtube) because it won't contaminate the prevailing mood which I want the main service to recommend.
- notyourwork 5y ago
  That's a good point and the next evolution recommenders try to do is to become contextual. There may be songs you like at dinner or on weekends but not vice versa. These types of things like time of day, device and other types of data points can make a recommender more precise.
soylentnewsorg 5y ago
You are exhibiting google's mistake, caused by over educated people with zero real experience thinking they know better. Or, I should correct that - some code they wrote knows better than the complex brain of the actual person. This results in "AI" (big if/then loop) trying to figure it out, and completely blocking any input from the customer. Here's the issue with "inferred from other activity." You know how you can find out? Give the customer a way to tell you. And no - you cannot "infer from other activity." >Skipping a song after less then 20 seconds for example can train a model... ..to be wrong. You know when I usually skip a song after 20 seconds? When I've heard it too much, or just heard it on youtube on my laptop an hour ago. In a couple of days I'll want to hear it again. The reason I heard it too much, is I like the song. Your AI though - if I skip this favorite song of mine a couple of times, will now block it and all other songs like it. The literal opposite of what I want. What this results in, is you have an inferior product, your customers leave, then you leave and go screw up another company's product without ever learning your lesson.
- djhn 5y ago
  I don't disagree with you but that last bit was quite unnecessary.
  - soylentnewsorg 5y ago
    That last part is the most important point - the main point. Your comment shows the exact issue as the "you don't need a dislike button because our AI knows better." You read my post, you read all the comments on here with users either jumping through hoops or ditching the product, you read the comment like what I replied to saying "we know what you need better than you do." That last bit is a person thinking some crap code they wrote can figure out why someone doesn't want to listen to a song at a particular time. They ignore all the inputs into that decision - how tired they are, what day they had, are they angry at something, have they heard it on another medium recently. Then ignore all that data they don't have, and have the AI make the decision: did you skip after 20 seconds? If you did, let's cut this genre out of your playlist. This little thing, destroyed spotify's recommendation product. The people in charge of this, are going to destroy everything they touch, because they think their little piece of code, is smarter than the user, about what the user wants. You know, because the user can't code. People making these decisions - well all I can say is I've met them. I've been personal friends with some. They have been passed up for jobs when I get called for a reference check. Because once they're on your team, they'll destroy your product, so they can feel personally superior. So that last part is very necessary, because it's an extremely serious issue, exactly as I put it. But in your opinion it's not, so let's cut that part out because it's not necessary for what I was saying. Because you know better than me what I was trying to say. <dislike>
laserlight 5y ago
> we’ve found from online experimentation customers generally don’t want to tell you they dislike something. I'm curious. Can you tell what kind of experiments you ran and what results you got?
solarkraft 5y ago
> Though some customers have shown desire to want to manage these signals Let me be one of those. I want to be able to control the algorithm in all of its parameters, but even the ability to control the values for new/familiar and how far to stray from the current genre would be very welcome.
syshum 5y ago
>>>one other consideration is disliking can usually be inferred from other activity. Skipping a song after less then 20 seconds for example can train a model GOD NOOOOOO that would be terrible, there are all kinds of time where I will skip a song because I do not want to Listen to it RIGHT NOW, not because I do not like it or never want to listen to it ever again... Music is very emotional, some times you just do not want to listen to X at the moment, 30 mins later you might.
newsbinator 5y ago
Yesterday I tried to book a plane ticket on eDreams and no matter which card/bank I used I got a bank authorization error. So I tried to contact eDreams, which requires a trip reference number. Since I couldn't pay them for a plane ticket, I didn't have a trip reference number. So all their ticketing and phone tree systems would hang up on me. eDreams customers, evidently, don't want to contact them without having a trip reference number. And they know this for sure because they've never had a customer contact them without one.
marcinzm 5y ago
The dislike button is if nothing else a sanity check for someone to have control if your model goes haywire. Realistically it will go haywire in a non-trivial number of cases. If every month your model goes haywire for 1% of users then you may see dislike used 1% of the time. That's tiny, right? However without it those users may leave and then next month another 1% will be hit and so on. Users remember the bad things a lot more than the good things so they will remember your AI f-ups and the lack of a way to fix it. The problem with data driven UX is that the data is often short term (ab tests are expensive) and doesn't cover long term complex user impact. Edit: For example, I’ve seen a lot of Etsy users complain about their front page being filled with NSFW recommendations and no way to get rid of them. I doubt they’ll be using Etsy much in the future.
greggman3 5y ago
This and your responses below remind can be applied to handicapped parking or handicapped ramps (ie, they aren't used by most users therefore we shouldn't have them). I prefer and use the dislike button when it exists. I don't care that 51% or 99% of users don't use it. Further, I want to send a concrete signal. Skipping a song doesn't mean I don't like it. It might just be that I didn't want to hear that particular song right now. Even if I skip it 10 times in a row that doesn't mean I don't like it.
yuliyp 5y ago
It's honestly really stressful using systems which do this. I have to worry about the signals I'm sending to the system all the time. Did I hover over that clickbait too long? Did I alt-tab while a song I liked was playing? Did I accidentally click like on something? Did I misclick onto the wrong song? Sure I might enjoy some clickbaity content every once in a while, but that doesn't mean I want the system to show me more of it, even if that will lead to me watching more things.
oriolid 5y ago
> Skipping a song after less then 20 seconds for example can train a model as well as incorporating an explicit dislike button. So, basically if I dislike a song but want to give it a chance and then decide that you want to never hear it again, your industry thinks that I liked it. That explains a lot.
somebodythere 5y ago
Disliking something is scary; if I press the dislike button am I going to be cut off from that whole artist/genre/etc?
kortex 5y ago
> I work in this industry and actually we’ve found from online experimentation customers generally don’t want to tell you they dislike something. > Skipping a song after less then 20 seconds for example can train a model as well as incorporating an explicit dislike button. Really? Cause I constantly wish for granularity in my ability to dislike something. Insta-skip might mean I'm just not in the mood. Or it might be "never play goddamn clownstep in my neurofunk mix". Or it might be "I actually love that song but I reserve it for special occasions to get into a specific headspace and if I listen to it too much it loses spark". Same signal, three radically different valences. I just wish services didn't treat all their customers like simps. I'd easily pay double for the ability to like, specify "Tuesday is flying microtonal banana day" or "start switching to prog house around 16:00 because that's when I start to get really deep into code but also easily frazzled." I'd pay gobs of money for Pandora's ability to discover new music with the library of Youtube+Spotify.

marginalia_nu 5y ago

I had to keep deleting my Spotify account because my recommendations kept getting contaminated with music genres I didn't like that were impossible to get rid of. Listening to the wrong song two times could irrevocably ruin my recommendations. I would have very much liked a dislike button, or at least a "forget everything you think you know about me"-button. Now it ended up with me getting rid of Spotify, because of this and other frustrations.

notyourwork 5y ago
Without specifics I cannot guess as to your specific case but in general customer's just do not use dislike buttons. So exposing them isn't a value add and usually a UI clutter that most customer's have no use for and won't use. In general, you had to be doing something to be getting recommendations so there is some action that when most customer's do demonstrates a high probability that you will like some genre and it just so happens you don't.
- chillingeffect 5y ago
  What industry just out of curiosity? I see dislikes used judiciously in youtube and reddit fwiw
  - notyourwork 5y ago
    Music but I don't work for Spotify currently.
- marginalia_nu 5y ago
  > in general customer's just do not use dislike buttons. So exposing them isn't a value add and usually a UI clutter that most customer's have no use for and won't use. A feature does not need to be used often to be useful. Take a fire extinguisher for example. Going by this logic, since most people will never even touch a fire extinguisher and most of them are never actually used, we should get rid of them because they're a cause of clutter.
- feanaro 5y ago
  I must say your dismissal of this is rather annoying. Multiple users are telling you they are facing a real -- and completely obvious and predictable -- problem due to the feature not being available. In a comment section regarding research showing it would be an improvement to recommendations. Increased UI clutter due to a single extra icon is not a good nor sufficient argument for not including it.
  - notyourwork 5y ago
    My dismissal is based on a many-million digit customer base and online real world data. Sorry if a few hackernews users feel they want a dislike button, most users don't and if you give to them most users don't use it. I don't mind disagreement but dismissing real world data from this industry is sort of silly to counter with 'because I want it you're wrong'.
    - feanaro 5y ago
      I'm not even saying you're wrong, I'm saying your argument for not including it for (at the very least) a sizeable minority is weak and somewhat silly.
      - notyourwork 5y ago
        What is a sizable minority aside from an oxymoron?
        hulahoof 5y ago
        A sizeable minority is something that is still visible on a pie chart.
        feanaro 5y ago
        A minority is anything that is not a majority (this is the definition I'm using). So a < 50% group can hopefully still be viewed as non-negligible, even though it is not the majority.
    - 5y ago
- breakfastduck 5y ago
  I absolutely love it when people working on a product tell their customers that they don't need a feature they're all asking for and that would clearly improve the experience, just because not everyone uses it all the time.
  - simion314 5y ago
    > absolutely love it when people working on a product tell their customers that they don't need a feature they're all asking for and that would clearly improve the experience, just because not everyone uses it all the time. We need a name for this, personally I call it GNOME mentality but probably is bad name since most developers are not familiar with GNOME (though the file picker issue get on top HN a lot )
  - notyourwork 5y ago
    > they're all asking for All is the flaw in your comment. One person does not imply all and data speaks for itself.
    - marginalia_nu 5y ago
      The data does not speak at all. You are drawing inferences based on the data, inferences of a type that are pretty easy to poke holes in by looking at the pretty large class of things rarely used but situationaly useful (often life saving). This type of bizarre logic makes me feel like I have to habitually use features even when they are not useful in order to protect them from UX designers in case I'm subject to some misguided A/B test.
      - breakfastduck 5y ago
        UX designers not sufficiently held in check by strong product teams are lethal.
      - notyourwork 5y ago
        > The data does not speak at all. What data are you referencing? I'm drawing conclusions based on data I have but I don't think you work for the same company as me so I'm not sure how you can speak to confidently about how wrong I am.
- toofy 5y ago
  Each of the comments that mirror yours illustrate a major problem we have in the tech sphere—assuming _we_ know better what the end user needs therefor we don’t listen to the actual user. In our defense, we’re very disconnected from the end users. I’ve done some work with a couple disaster organizations in the past and we continually get massive praise and thanks from the victims of these disasters. The actual victims have the same praise over and over again: Unlike the red cross, salvation army, and FEMA we actually ask them what they (the victims) need. We don’t assume they need water and and a tarp, we assume they know whether or not they already have those things. Often the victims need diapers, help with flood mold remidiation, a chainsaw to cut limbs from their street so they can get their cars out so they can go help the neighbors a few streets over etc…. They know what is prohibiting them from getting out to help others much better than we. When people are telling us what they need, we should listen to them and we should actively try to avoid our tendencies towards “I know better what you need.” I have a couple friends who work on recommendation algorithms full time and in their defense, they’re fully aware how terrible the recommendations are—I still get my music recommendations from friends, from the local music store employees, and friends who work in music venues—their recommendations are infinitely better.
the_only_law 5y ago
> Listening to the wrong song two times could irrevocably ruin my recommendations. Similarly any app or website that pulls “top artists” from Spotify tends to include stuff I listens to a bit several years ago, but never stuff I’ve listened to nonstop for months.
0134340 5y ago
Yes that's very frustrating. I'm scared to even listen to new music on it because I'd try to use it for discovery but it would mistake what I listen to for what I like. Besides that it hardly recommends me any new music anyway. I now just use youtube for discovery.
- marginalia_nu 5y ago
  They should add an incognito mode or something. You could even do this maliciously. Send someone a Spotify link to a Steely Dan song and a Skrillex song, and now yacht rock and dubstep will follow them around until they delete their account (I'm assuming the intersection between those fan-bases is vanishingly small).
  - Oddskar 5y ago
    That already exists. Profile > Private Session
    - marginalia_nu 5y ago
      Doesn't that just hide your activity from other people?
      - Oddskar 5y ago
        From their support docs: Note: Anything you listen to in a Private Session may not influence your music recommendations, e.g. Discover Weekly.
  - vanilla_nut 5y ago
    High school me would have been very offended to be labeled “vanishingly small.” But you’re not wrong. By the way, I believe Spotify does have a private mode, which at least blocks your plays from showing up in the social pane and on your profile. But I’m not sure if it cuts them out of your recommendations. Please, dear God, Spotify: just because I listened to the Hamilton soundtrack once does not mean that I want to hear it every week in Discover Weekly and multiple times per Duo playlist, because Lin-Manuel Miranda is apparently a genre of himself.
1_player 5y ago
I keep hearing from people how good is Spotify "Discovery" playlist. I've been a paid customer for 5+ years and their recommendations have a hit rate of 2%. I listen to a lot of different genres, artists, mostly dictated by my mood. Today it was some easy Dire Straits song, yesterday QOTSA, the other day some psytrance. So Spotify recommends all genres in the same playlist, creating an awful cacophony of songs that don't flow into one another. Spotify recommendation system is very bad in cases like mine apparently. It's so bad I keep listening to the same old songs instead of discovering something new, which is why I'm no longer a customer until they improve their AI.
- breakfastduck 5y ago
  I listen to a LOT of different genres of music and their Discover reccomendations (not the playlists, the actual you like X artist you might like X) are absolutely brilliant. Got to be approx 50-60% hit rate. It seems to differ wildly between people - I've heard plenty of people making the same case as you, but many that have the same as me. It never appears to be 'its ok', its either 'brilliant' or 'terrible'. I've been using Spotify for 10+ years at this point, so maybe that makes a difference?
- newsbinator 5y ago
  Same here. I like Christopher Tin (e.g. Baba Yetu) and therefore for 6 years I've had no way to tell Spotify I don't like 8-bit video game music. No amount of skipping or "I don't like this artist" seems to convince Spotify to stop offering me variations of 8-bit Tetris themes or MULTIPLE 8-bit Cowboy Bebop Tank! theme variations per every single Discover Weekly. I'd be thrilled to even get a 2% hit rate on music I like. Seriously Spotify: two separate "Tank!" versions per Discover Weekly is too many.
  - kortex 5y ago
    Man, the spotify algo makes no sense. The other week I was in a really specific mood, starting with Sonic 1/2 OST and moving into 8bit/chiptune/keygen. I probably listened to every rendition of Chemical Zone out there. Rest of the week, much the same. Nothing in Discover Weekly.

user-the-name 5y ago

> one other consideration is disliking can usually be inferred from other activity. Skipping a song after less then 20 seconds for example can train a model as well as incorporating an explicit dislike button. This is nonsense. The fact that I skip a song does not mean I don't like it in general, it means I don't want to hear it right now. Music taste changes with mood and situation. If you treat a skip as "I don't ever want to hear this again" you will collect absolute garbage data, and if I know you are doing this, it will make your app unusable because I can't do normal everyday actions without ruining my recommendations.

notyourwork 5y ago
It depends right. If you skip a song everytime its played in a sequence is a stronger indication of dislike compared to skip it once but listening to it a few times. The point is there is more to context and recommenders than a adding a dislike button and having a magically awesome recommender.
- oriolid 5y ago
  Have you read about illusion of control (search for elevator door button for one well known example)? If we take the idea that recommendations are really hard and dislike button doesn't help at face value, wouldn't it still improve the user experience to give them the illusion that they can directly affect the recommendations? I'm not really convinced about recommendations being hard, though. Pandora could do them a long time ago, Netflix had good recommendations at one point but not since they dropped the star ratings and at some point Amazon wasn't that bad either.

rednalexa 5y ago

Is this within a global context or a local one? Liking a song in Spotify is global - Spotify then adds it to a list of liked songs. But from my experience with Pandora, radio stations were localized. Liking a song did add it to a list somewhere, yes, but disliking a song just meant that a particular radio station on Pandora wouldn't have a song in that particular style, and other stations did not seem to be affected by this. In other words, what you want to listen to, for me, seemed to be very contextual. And since I felt my dislikes were captured in a context (radio station), I didn't have any qualms disliking frequently using Pandora. And my experience with Pandora for song discovery has always been much better.

ozzythecat 5y ago

> Skipping a song after less then 20 seconds for example can train a model as well as incorporating an explicit dislike button. How do you know I dislike the song? Instead, maybe I just heard it (or a similar song) on your platform or elsewhere outside of your app and don’t want to listen to it again right now? I guess you could try to build a more complex model that determines what songs or music I like at what times of the day or what day of the week, or based on what I just listened to. IMO this is rampant in FAANG companies. We create complex solutions to problems that didn’t really exist and pat ourselves on the back when our experiment show positive results. The experiment might show you some metric (total time in app, play time at whatever percentile, etc) increased. Great! Now we launch the model in production for all customers and pat ourselves on the back. Meanwhile, the actual experience became shittier. A few promotions come out of this, and we continue living in the bubble. Alright, end of rant.

notyourwork 5y ago
The paper is making an obvious claim. If I have data showing positive and negative affinity, that will perform better than a model which just has positive day. I don't find this very surprising but my point is in practice exposing dislike buttons to customers are a CX that is hardly ever used. So the point that a model will perform better with data that customer's rarely offer up to your platform makes their claim moot in my opinion.
- corin_ 5y ago
  I downvoted your comment to demonstrate that I disagree with you. I've heard some good arguments against allowing users to show dislikes/downvotes as well as likes, but I've not heard your argument before and it doesn't strike me as accurate. Less used than the positive direction, sure, but still significantly used when available.
  - notyourwork 5y ago
    Netflix, Youtube and reddit aren't the same medium as short lived music streaming is. Streaming music is a show lived precision game of guessing what's relevant. There is a lot of it depends scenarios, are we showing recommendations on a landing page, are we sequencing songs in a station or are we ordering results on a search page. The CX context is a big part of how you can use signals to benefit the customer. My argument isn't an argument so much as a from real world data in this industry at a major industry competitor that customer's don't use the dislike button enough to justify it taking over your CX. Using co-occurrences of playback across a customer segment and combining that with likes or other signals is good enough.

Groxx 5y ago

Skipping doesn't imply dislike, though it does correlate (because the reverse is generally true, people skip songs they dislike). But I routinely see skip mentioned as non-downvote sources of dislike information. Easily half of the times I skip it's because I don't want it right now. Some of my favorite songs reproduce absolutely horribly in a car on a highway for example, because the surrounding noise utterly destroys subtleties. So maybe you add "but it's not a downvote if it's hearted or otherwise positively ranked"... but many other times I'm just not in the mood for [a song I had never encountered until now].

anticensor 5y ago
Correlation does not imply causation, and neither does causation imply correlation.

ics 5y ago

A problem I have with the inferred dislike is that when I skip a song it's only something I dislike half the time. The other half I'm simply not in the mood to hear it, maybe for another 10 minutes or another 10 months. The result has been that I've become more willing to listen to certain genres on another platform (say Youtube) because it won't contaminate the prevailing mood which I want the main service to recommend.

notyourwork 5y ago
That's a good point and the next evolution recommenders try to do is to become contextual. There may be songs you like at dinner or on weekends but not vice versa. These types of things like time of day, device and other types of data points can make a recommender more precise.

soylentnewsorg 5y ago

You are exhibiting google's mistake, caused by over educated people with zero real experience thinking they know better. Or, I should correct that - some code they wrote knows better than the complex brain of the actual person. This results in "AI" (big if/then loop) trying to figure it out, and completely blocking any input from the customer. Here's the issue with "inferred from other activity." You know how you can find out? Give the customer a way to tell you. And no - you cannot "infer from other activity." >Skipping a song after less then 20 seconds for example can train a model... ..to be wrong. You know when I usually skip a song after 20 seconds? When I've heard it too much, or just heard it on youtube on my laptop an hour ago. In a couple of days I'll want to hear it again. The reason I heard it too much, is I like the song. Your AI though - if I skip this favorite song of mine a couple of times, will now block it and all other songs like it. The literal opposite of what I want. What this results in, is you have an inferior product, your customers leave, then you leave and go screw up another company's product without ever learning your lesson.

djhn 5y ago
I don't disagree with you but that last bit was quite unnecessary.
- soylentnewsorg 5y ago
  That last part is the most important point - the main point. Your comment shows the exact issue as the "you don't need a dislike button because our AI knows better." You read my post, you read all the comments on here with users either jumping through hoops or ditching the product, you read the comment like what I replied to saying "we know what you need better than you do." That last bit is a person thinking some crap code they wrote can figure out why someone doesn't want to listen to a song at a particular time. They ignore all the inputs into that decision - how tired they are, what day they had, are they angry at something, have they heard it on another medium recently. Then ignore all that data they don't have, and have the AI make the decision: did you skip after 20 seconds? If you did, let's cut this genre out of your playlist. This little thing, destroyed spotify's recommendation product. The people in charge of this, are going to destroy everything they touch, because they think their little piece of code, is smarter than the user, about what the user wants. You know, because the user can't code. People making these decisions - well all I can say is I've met them. I've been personal friends with some. They have been passed up for jobs when I get called for a reference check. Because once they're on your team, they'll destroy your product, so they can feel personally superior. So that last part is very necessary, because it's an extremely serious issue, exactly as I put it. But in your opinion it's not, so let's cut that part out because it's not necessary for what I was saying. Because you know better than me what I was trying to say. <dislike>

laserlight 5y ago

> we’ve found from online experimentation customers generally don’t want to tell you they dislike something. I'm curious. Can you tell what kind of experiments you ran and what results you got?

solarkraft 5y ago

> Though some customers have shown desire to want to manage these signals Let me be one of those. I want to be able to control the algorithm in all of its parameters, but even the ability to control the values for new/familiar and how far to stray from the current genre would be very welcome.

syshum 5y ago

>>>one other consideration is disliking can usually be inferred from other activity. Skipping a song after less then 20 seconds for example can train a model GOD NOOOOOO that would be terrible, there are all kinds of time where I will skip a song because I do not want to Listen to it RIGHT NOW, not because I do not like it or never want to listen to it ever again... Music is very emotional, some times you just do not want to listen to X at the moment, 30 mins later you might.

newsbinator 5y ago

Yesterday I tried to book a plane ticket on eDreams and no matter which card/bank I used I got a bank authorization error. So I tried to contact eDreams, which requires a trip reference number. Since I couldn't pay them for a plane ticket, I didn't have a trip reference number. So all their ticketing and phone tree systems would hang up on me. eDreams customers, evidently, don't want to contact them without having a trip reference number. And they know this for sure because they've never had a customer contact them without one.

marcinzm 5y ago

The dislike button is if nothing else a sanity check for someone to have control if your model goes haywire. Realistically it will go haywire in a non-trivial number of cases. If every month your model goes haywire for 1% of users then you may see dislike used 1% of the time. That's tiny, right? However without it those users may leave and then next month another 1% will be hit and so on. Users remember the bad things a lot more than the good things so they will remember your AI f-ups and the lack of a way to fix it. The problem with data driven UX is that the data is often short term (ab tests are expensive) and doesn't cover long term complex user impact. Edit: For example, I’ve seen a lot of Etsy users complain about their front page being filled with NSFW recommendations and no way to get rid of them. I doubt they’ll be using Etsy much in the future.

greggman3 5y ago

This and your responses below remind can be applied to handicapped parking or handicapped ramps (ie, they aren't used by most users therefore we shouldn't have them). I prefer and use the dislike button when it exists. I don't care that 51% or 99% of users don't use it. Further, I want to send a concrete signal. Skipping a song doesn't mean I don't like it. It might just be that I didn't want to hear that particular song right now. Even if I skip it 10 times in a row that doesn't mean I don't like it.

yuliyp 5y ago

It's honestly really stressful using systems which do this. I have to worry about the signals I'm sending to the system all the time. Did I hover over that clickbait too long? Did I alt-tab while a song I liked was playing? Did I accidentally click like on something? Did I misclick onto the wrong song? Sure I might enjoy some clickbaity content every once in a while, but that doesn't mean I want the system to show me more of it, even if that will lead to me watching more things.

oriolid 5y ago

> Skipping a song after less then 20 seconds for example can train a model as well as incorporating an explicit dislike button. So, basically if I dislike a song but want to give it a chance and then decide that you want to never hear it again, your industry thinks that I liked it. That explains a lot.

somebodythere 5y ago

Disliking something is scary; if I press the dislike button am I going to be cut off from that whole artist/genre/etc?

kortex 5y ago

> I work in this industry and actually we’ve found from online experimentation customers generally don’t want to tell you they dislike something. > Skipping a song after less then 20 seconds for example can train a model as well as incorporating an explicit dislike button. Really? Cause I constantly wish for granularity in my ability to dislike something. Insta-skip might mean I'm just not in the mood. Or it might be "never play goddamn clownstep in my neurofunk mix". Or it might be "I actually love that song but I reserve it for special occasions to get into a specific headspace and if I listen to it too much it loses spark". Same signal, three radically different valences. I just wish services didn't treat all their customers like simps. I'd easily pay double for the ability to like, specify "Tuesday is flying microtonal banana day" or "start switching to prog house around 16:00 because that's when I start to get really deep into code but also easily frazzled." I'd pay gobs of money for Pandora's ability to discover new music with the library of Youtube+Spotify.