Live data from Hacker News

Dislike button would improve Spotify's recommendations

news.cornell.edu

11–20 of 660 posts

Re: Dislike button would improve Spotify's recommendations

#11

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. Thou…

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…

> 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.

Re: Dislike button would improve Spotify's recommendations

#12

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. Thou…

> 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.

Re: Dislike button would improve Spotify's recommendations

#13

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. Thou…

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.

Re: Dislike button would improve Spotify's recommendations

#14

I'll never use Spotify again. My account was hacked by a latino and now my recommendations is 100% latin music. They told me there is nothing they can do to fix this. My recommendations are broken and my only option is to create another account, cancel my subscription, buy it again... Not gonna happen.

I found out you can break your recommendations without losing control of your account.

I listened to one podcast episode and it's been stuck on my home screen for two months. I've tried listening to it, ignoring it, unsubscribing, deleting (you can't!). It's just there forever now.

And of course, that means it's always recommending related podcasts all the time so most of my homepage is now filled with this stuff and I have no way to remove it.

Re: Dislike button would improve Spotify's recommendations

#15

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. Thou…

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.

Re: Dislike button would improve Spotify's recommendations

#17
post #8

i am actually impressed to see this come out cornell. i tend to see like and dislike buttons as too simplistic. they greatly reduce why a user would dislike something. i am annoyed at youtube only allowing me to choose from "i didn't like it" or "i have already watched it" for example. i have thousands of other reasons. that's why i don't use dislike buttons.

A dislike button is too simplistic, but having no way to dislike something is extremely bad user experience. How do you train a recommendation engine without a negative signal?

Sometimes I dislike something and I don't want it around anymore. Which is the reason I've migrated to Youtube Music, which is an inferior product, but has a dislike button.

Re: Dislike button would improve Spotify's recommendations

#18

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. Thou…

> 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.

Re: Dislike button would improve Spotify's recommendations

#19

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. Thou…

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].

Re: Dislike button would improve Spotify's recommendations

#20
post #15

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. Thou…

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 recomm…

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.
Post reply on HN