As a result, the signal coming from the dislike button was not very useful, and just added unnecessary complexity. Instead, the algorithm can present the user a fixed song in different settings to see if they continue to skip it. A nice result is that the user doesn’t have the negative reaction “hey I disliked this song why are you presenting it to me again”
This alone drove me back to my local music library.
I would rather dislike a song and have it recommended in a "How about now" playlist to determine if I actually hate it or if it was a mood in the moment.
I think the main issue here is, is that users feel a bit powerless to influence the algorithm with thier dislikes as much as thier likes. Even though a song skip may be functionally similar to the dislike button, the user has no "haptic feedback", for lack of better term, when they really feel the need to have their negative opinion on the song heard by the algorithm.
I think most people's gripes with the nature of these recommendation algorithms deciding things, centre around the feeling that the algorithm doesn't listen to them and their opinion as much as it obeys other overriding trends. Giving them tangible and tactile options to deliver their opinions in hard and fast way gives them some more peace of mind that the algorithm is working for them, and not for some other entity.
Some songs have extremely long shelf life and others are extremely short. For instance, an "adult contemporary" station could probably get away with playing Elton John's "Tiny Dancer", a 49-year old song, without anyone thinking anything is weird.
More popular songs of that year such as those by Roberta Flack or Gilbert O'Sullivan would not get the same non-reaction
Some songs have even longer shelf life that are associated with real world things like birthdays, holidays, seasons, etc. It's pretty common to hear a set of Christmas music that was literally recorded during WW2 with everyone's approval.
Some songs grow on people, others grow tired quickly. The classic example is novelty songs that specify a dance such as the cover of The Birdie Song by The Tweets or Black Lace's Agadoo, which reached impressive popularity and then were promptly forgotten forever. Ones that you may be familiar with such as Rick Martin's Macarena, probably gave you a headache just thinking about it.
But not all. Chubby Checker's "The Twist", a song in this category of instructional dance pop music, has escaped the ban hammer of time.
But overall the pattern of memory holing proscribed dance songs is robust, even for songs that developed a dance without actually calling it out, such as "Achy Beaky Heart".
Also there's no asymptotic drop after release date. Some songs that are popular now and symbolic of an era weren't as popular when they were new and some have second lives when featured in say a television show, video game, or covered by a more contemporary artist. Every few years a new version of Gershwin's "Summertime" seems to come out followed by renewed interest in previous versions by people like Billie Holiday, Janis Joplin and Ella Fitzgerald. Should Ella Fitzgerald albums be recommended to people for the next 18 months now? Probably not the right way to interpret those results.
There's an exogenous contextual reason for such things and those have to be accommodated for
It's completely non-trivial and a simple model of skips and likes without sophistication behind it will dramatically fail to be anything other than an irritation
It was done by a group from Seville. Los del río. Very southern Spanish accent which is quite different from Ricky Martin.
I'll need to review 90s Latin pop again. Sorry about that.
I read that YouTube Music recommends different music depending on time of day, whether I'm at work / home / in the car, and even takes the weather into account. I'd be surprised if Spotify didn't have some context awareness as well.
I recently got super into the band Squid. But I hated them the first time I heard them. It was because I was depressed and looking for more music like Black Midi-- Black Country New Road came up and was perfect. Squid came up and made me want to punch my speakers.
If they have enough data to predict my mood to that extent then I would like to hire them as my therapist.
Ideally, they should let me like, dislike, skip, hate, completely block artists, and allow me to like multiple times, to teach them the ones I like over and over from the ones I like a little.
So basically Spotify is performing reinforcement learning on you— beautiful!
If they let you block artists then this is less money for them.
I don't know why these show up in my recommendations every single day. I will never watch them.
There is a way to do this, the signal is adding the song to your favourites or a playlist
Before, I had to skip the songs that I did not like, and Deezer was keep putting them back in my 'Flow' a few days later. That was pretty annoying.
>To dislike a song is easy – to like one, you have to actually invest time in it.
Likewise, to detect if a user dislikes a song is easy since the user will not invest time listening to it: the skip button is the dislike button.
In contrast, detecting if a user likes a song is more difficult: there needs to be a way to distinguish between a user enjoying a song and the user not paying attention (in that case the user will provide no input whether they like the song or not.)