Pretty cool to see them go after the cold start problem with audio models. Often measuring the impact of having cold start vs. not is very difficult because it only affects the long tail of tracks at first, but eventually it could change the dynamics of discovery in the system.
It occurred to me, though, that this is wide open to hackery.
Find a popular song that is likely to be on a lot of people's playlists. Make a new song that is a close match to it for raw audio modelling. Launch on Spotify.
Yeah you'll probably only get one or two plays per user, but spread over a large number of users (who all listen to the whole Discover playlist every Monday, like I do) it's still significant traffic.
I imagine that once you get to any significant number of plays (N=1000, which would not be a significant payout on spotify at all), the collaborative filter model would take over and if none of the people who listened to it caused further positive signals as mentioned (visiting the artist page, repeat listens) then the track would stop being recommended pretty quickly.
Also, what you are describing sounds like what producers already do because people actually like it: find the latest trends in sound and copy them, (but hopefully with a fresh twist so that people get into it) :)
It's also how Criticker recommends films, and I have to say their predicted scores are uncanny. After watching a film, I often think of a 1-100 score and then look it up on their site, and it's at most ±2 from it.
Google Music's "I'm feeling lucky" is awesome. This is very subjective, but I feel it makes much better predictions as to what I'll like than Spotify. Maybe someone else has a different experience?
Google Music was, in my opinion, the industry's greatest music discovery service back when it used to have the "Explore" tab where you could drill down through genres and sub-genres to find popular or classic albums within that genre. It also used to give me much better recommendations on the homepage for new albums that I might like.
It's all gone down hill since they switched everything to mood-based radios. I feel like the main problem is that I like to listen to albums and Google seems to assume I only want to listen to random streams of disconnected singles. I seem to just get played the same stuff over and over, too.
I don't know how to find new (new to me, not the world) music anymore. Are there any good services that anyone recommends?
> I don't know how to find new (new to me, not the world) music anymore. Are there any good services that anyone recommends?
I have the same problem. Usually I find new stuff on the various music subreddits and last.fm's Similar Artists page. Even though I have tons of playlists and saved songs on Spotify, their recommendations usually suck (probably because I listen to just about every music genre under the sun.)
Their radio feature rehashes the same songs I've heard dozens of times, and Discover Weekly is usually so awful I never listen to it.
I disagree, Last.fm's recommendations have always felt lukewarm to me. Discover Weekly, on the other hand, has introduced me to TONS of new music that I still regularly listen to.
In my experience they got pretty close to Pandora's recommendations already, which were excellent.
last.fm for me just recommended the next big artist in the same category. Often their music styles were still obviously different and the recommendation rather put me off. People that are hooked by the complexity, details and perfectionism in "Nightfall in Middle-Earth" won't necessarily like Manowar.
I recently got one Spotify recommendation that lead me into listening through a band's full catalogue and getting tickets for their show two weeks later (Insomnium, btw). It also dug out a song that I liked in primary school but completely forgot about. They discovered that I like cheezy metal covers of 80s pop songs and add one by some obscure band to my list every now and then. I'd say my experience was often pretty accurate.
Thanks, that is one of my main gripes with most recommendation engines, and since you like metal, the examples really resonated with me. For me, it was Wardruna - an ambienty, folky acoustic group. Listeners often also like the "typical" viking metal bands, but damn, if I want something like Wardruna, I do not want death metal singing about the same topics, I want, say, Forndom. The same often happened in electronic music as well: if I listened to 5 relatively unknown psytrance songs I liked because they have a nice balance of melodies layered with very distinct synth sounds, then no, I don't want to liten to David Guetta next, because that is also trance-like.
In other words: yes, last.fm's recommendation was much too much based on "customers also liked", which helps in a lot of cases, but so often it horribly fails ("Customers who bought The Martian on Blu-Ray also bought this asthma medication because chance happens. Wanna try it out?").
Last.fm's customized radio didn't work well for me - it'd play the same limited selection of tracks every time, often one song from an album and never play any of the other ones on it.
> Unlike Netflix, though, Spotify doesn’t have those stars with which users rate their music. Instead, Spotify’s data is implicit feedback
I wish they kept it simpler with ratings, and explicit instead of implicit. The discovery weekly playlists are absolutely horrible for me to the point that I don't even bother checking them nowadays.
What works better for me for discovering new music with spotify is right clicking on an existing playlist and then "Create Similar Playlist" - that gives way more control over what kind of genre/style should the playlist consist of.
Explicit and implicit rating systems will never fit everyone.
But I will go as far and say that implicit rating fits a lot more than explicit, because it doesn't require the user to do extra work on top of the base goal of listing to "good" music.
I feel like explicit rating scales very poorly with catalogue size. So when you have music, and as much music as Spotify has, then the work effort of explicitly rating your taste becomes too big, you start to not bother, and quality of rating becomes poor.
There's no reason that I can see not to combine both explicit and implicit rating. If I really love a song let me mark it as such, but also feed things that I've chosen to listen to/skip/add to my library into that algorithm (possibly with a lower weight).
I am still missing Rdio in that regard, they were just playing the right songs all the time. With Spotify this is absolutely not the case and it's even far worse than YouTube.
While I haven't found a recommendation engine that works well enough for me, rdio's was very close to it.
It's big feature, among others, was their heavy rotation section - I only followed people that shared a similar taste to mine and we had a really cozy circle of listeners whose current favorites were surfaced by said heavy rotation section. We discussed albums in comment sections, shared playlists and I regularly stumbled upon familiar usernames and friends when discovering new gems.
This social component to discovery is completely missing from Spotify, yet it's more powerful than any recommendation engine I have used, including Spotify's attempts, which I'd rate as mediocre.
I use Spotify's Discover Weekly on a regular basis. On average I like more than half of the songs on the playlist, and it's my main way of discovering new music.
Discover Weekly works really well for me. It surfaces a lot of music I've forgotten about and for the most part provides an interesting collection to listen to. Having said that, I have a very eclectic taste in music, so it's probably harder to hit on things I won't like. My process is to try to listen to it several times through and then pick out the stand out tracks once everything has had a chance to grow on me.
One thing I worry about with it though is how much my behaviour might influence the choices. For example, if there's a track in there that I already know quite well, because I like it, I'm often scared of skipping it in case the algorithm takes that as a massive negative signal.
Thinking about it a little more, we've listened to a lot of Moana / Frozen / Disney on my Spotify account and none of that has come up in the Discover Weekly, so I guess they filter for it somehow already. Though I guess when they compare to other people with the same tracks you'll probably get a bunch of listeners that fall into a similar category anyway. Also, it doesn't sound like the number of times you play a track adds a huge amount of weight anyway.
Also, I have a young daughter who often asks me to play music in the car or Google Home and those affect my recommendations. I wish Spotify had some switch I could turn on to temporarily ignore anything I did until I switched it back off.
In terms of behavior influence, I listened to pretty much exclusively Hamilton for like a month and didn't get very many new Broadway or hip-hop songs (which is good for me because I tend to not really listen to either genre). So there must be significant weight given to songs/genres you've historically listened to a lot.
For anyone interested in the Raw Audio Models section of that articles, there are some fun endpoints[1][2] from the Echonest API that provide those models.
They've moved over to the Spotify API since I last had a play, but it's great that they still provide them.
You can get the audio breakdown of a track, as well as a summary of the track features including fun stuff like "danceability" and musical positiveness ("valence"). Radiohead's "Fitter, Happier" was low on both of these points if I remember correctly.
I also wanted to nitpick this part of the article -- the neural network isn't determining the audio features; these are all hand-engineered features developed by The Echo Nest.
The neural network is trained to mimic collaborative filtering vectors from raw audio. It's a separate model from time signature, key, mode, tempo, loudness, etc.
They still suck compared to what Last.fm recommends me. Also the fact that Spotify uses implicit feedback instead of explicit like Last.fm pisses me off. I'm constantly left asking what the algo would make of my actions.
Spotify's discovery engine playlists (Discover Weekly and Release Radar) just don't work for me. I now ignore both playlists and I suspect that the lists are influenced by payola.
I've tried training the algo by following artists and saving albums in the style that I would like, but these playlists keep peddling stuff that is way off the mark. Interestingly, the daily mix playlists have responded to this training, but not Discover Weekly or Release Radar.
For users who are tired of the same songs being fed to these playlists each week, you can create an IFTTT action to save the content of these playlists in separate archive playlists. Once a song is in the archive, it won't (or shouldn't) appear in either of the weekly so-called discovery playlists.
> Once a song is in the archive, it won't (or shouldn't) appear in either of the weekly so-called discovery playlists.
They still do, sometimes, at least for me. But more importantly, adding songs to playlists sends the positive signal to Spotify’s recommendation engine, as far as I know. So, saving all your discover weekly tracks into archive playlists will encourage Spotify to give you similar music in the future.
Discover Weekly works well enough for me - I do wish I could exclude devices from influencing that playlist as what I listen while gaming on PS4 is completely different from the rest of the day. It also seems to have a bias towards “big” commercial releases.
Spotify’s “intelligence” in general is a huge let down though. Radio stations are extremely limited - more like 15-song static playlists indefinitely on repeat! It annoys me to no end, same for daily mixes. I end up listening to the same songs over and over and over and over again. Maybe that’s what makes them the most money?
Last.fm was amazing at finding me new music I liked. Rdio was amazing at.. radio :D I used to go for months on the same station. I miss both a lot, and occasionally I still use last.fm or everynoise.com to generate better playlists for Spotify.
I don't listen to pop very often - that's how I notice it. Eventually I'll realize that a song it added to my playlist is playing in the background in a shop, on TV or the radio.
Spotify's daily mixes seperate the genres/moods. I occasionaly switch to listening electronic music, but I mostly listen to rock. Spotify has one daily mix of rock, one daily mix of heavy rock and one daily mix of electronic. It doesn't mix the genres.
But discover weekly only gives recomendations for the dominant genre.
With all that AI excitement Spotify, Google Play Music and Apple Music misserably fail at generating good musical recomendations. At least for me. Spotify Discover Weekly was no different for me when I still used Spotify (six months ago).
Not only it is not personalized at all it is also quite lame. I am doing enough to seed - following enough artists, genres, liking album, subscribing to playlists but the quality of recomendations is mediocre at least.
For example at the moment Apple Music is suggesting to me four Wednesday playlist - all heave metal the genre I never liked and listened to. Also two artist spotlights - Jeff Chang and Danny Chan. Yes I have Japanesese account but this is Cantopop and Taiwanese crooner - so a little off the map especially that I do not enjoy pop music at all especially Asian. Some other examples are what I call comin denominators. Yes I follow lots of jazz but Frank Sinatra is not jazz music, nor Tony Benett. etc...
My experience was that Spotify radio is utterly bad, Discover Weekly quite okay, but the weekly Apple Music playlist really nails it. I found a lot so far that I didn't know before and added to my library.
Same here with Netflix: their recommendations have been a complete failure. It seems like Netflix is pushing mostly their Netflix original series. What does works for me in Netflix are the similar to this below the movie, e.g. after you've viewed the movie. But that is a very simple algorithm any sophomore could write, not related to the recent developments in neural nets/machine learning.
When I go to an filmfestival or an concert, part of what I pay goes to the proffessionals who make the selection for me. And I'm happy to pay their services, just like I pay for journalists.
I have the complete opposite experience compared to you. All of my Discover Weekly and Daily Mixes are at least 75% music I like, even if I didn't know it beforehand, or only knew the artist by name.
At bare minimum I really with I could tell these services which artists I explicitly don't like so that they're never recommended to me. That alone would be a huge improvement.
Disliking multiple songs from the same artist only to have it suggest the next one from their repertoire is infuriating. Worse, Apple Music will play songs you've disliked in the radio anyway.
Discover Weekly is weird. For the longest time it kept recommending Finish music (roughly 30% of all recommendations), I'm from Sweden and barely ever listen from anything from Finland (or Sweden for that matter).
They have a ridiculous bias towards covers. I bet I've been recommended 50 (I wish that was an exaggeration) version of the Gladiator theme ('now we are free'). And they are all terribly bad (as in blood coming out of my ears bad). I am genuinely ashamed that someone thought that it would be a good idea to submit it to spotify - and if so how spotify could recommend it to anyone, there is no way spotify could have gotten any indication that anyone has ever liked any of those versions of that song. Similarly I get lots of game of throne covers, and now Despacito covers (never listened to that song in any variant on spotify, on purpose at least).
I have a similar problem with remixes from trance tracks I listen to regularly. Some of them have literally 30 different remixes, and they keep getting into the 'discover weekly' playlist. I guess these would be coming from the acoustic profile or the NLP recommendation algorithm described in the article. That said, I still like the feature a lot, it's a little naive to think it could only ever recommend songs you will actually enjoy.
A curious side effect of Discover Weekly is that it sometimes influences what I listen to, i.e. I'm afraid to listen to a song in a particular genre because I don't want Disover Weekly to get the wrong idea. But then again, maybe Discover Weekly knows me better than I do.
I had the same issue: I spent a couple of days listening to old tunes I rarely pick anymore, and it “messed up” my Discover Weekly. I thought “Well I only have myself to blame, don’t I?”.
Turns out Spotify has a “Private session” mode to prevent influencing your taste.
Just listen to music. Don't worry, listen to what you like, create your own playlists. Use Discover Weekly for inspiration, not your sole source of music.
Sometimes I wish Spotify added a bit more 'noise' to their recommendations, so to speak. If I don't listen to much music except Discover Weekly for a few weeks, I (subjectively) find that what's recommended to me more or less sounds the same after a while. Either they are afraid to insert new things that stray too far from an optimal recommendation or they forget too much of my listening history.
I'm afraid it is so extreme that I feel lots of my discover weekly recommendations are based on solely of what I listened on discover weekly last week.
Which would be quite bad, but that is how it often feels.
I have the same problem. And we it every fad machine learning thing. It never works for me but a lot of people are happy. Does anyone want me as a test subject?!
I’ve noticed that there are certain seemingly-esoteric songs that many people are assigned, and subsequently feel let down upon realizing that they’re (currently) common assignations. It wasn’t because the algorithms recognized their excellent personal taste.
“Didn’t I” by Darondo seems to be a good example of this. A fantastic, forgotten soul song (and from the Bay Area!) that was assumed to be a recognition of my friend’s unique taste in his car... until the other four people inside revealed they’d all had it in the last two weeks, too.
It makes me wonder about the motivations. Did the IP for this song recently change hands?
(I still get the shivers recalling the month when people kept gleefully playing me “Temporary Secretary” by Paul McCartney; a new discovery from their Discover Playlists. I could have marched to Stockholm to strangle a data scientist.)
>“Didn’t I” by Darondo seems to be a good example of this. A fantastic, forgotten soul song (and from the Bay Area!) that was assumed to be a recognition of my friend’s unique taste in his car... until the other four people inside revealed they’d all had it in the last two weeks, too.
It could be popularized by DJs who are playing it, e.g. Four Tet, whose personal playlist on Spotify now has close to 39 thousand followers.
At least in that case specifically, I doubt it's due to distribution rights. There are no new releases indicated on Discogs, which seems likely if someone new acquired them.
It was also used in “Breaking Bad”, in the scene where Walter sets “Ken Wins” BMW on fire. Thomas Golubić was the music supervisor for BB and Better Call Saul and these shows’ playlists make superb additions for Spotify listening, in my opinion. They’re chock-full of great discoveries like Darondo.
It doesn’t surprise me that “Didn't I” comes up in many playlists, but it does surprise me that the other great songs on that album don’t get worked into Spotify stations. This is a great example of their algorithmic problems.
Rdio had an option to address that: Its radio / "play similar stuff"-feature would let you select how far you wanted to stray from the original source of the generated playlist. If I remember correctly it went from "same artist" to "adventurous" in 5 steps.
Tell me about it. A large part of the community we had at rdio used to meet at in a post-rdio slack channel after the service went down. For months we discussed alternatives and how mediocre and barebones they were compared to rdio. Most of them, including me, settled for Spotify, but the community is pretty much gone.
On the other it led me to double down on my purchased offline music collection, which I already maintained long before rdio.
Yup, I worked on the playlisting service for a little while at Rdio, it was pretty simple but worked well. Based on those presets we would decide how far along the artist similarity graph we would walk from the original playlist "seeds".
For some reason every week I get like 2-5 songs with the same beat, but a bit remixed or covered by someone else. It's trying way too hard to recommend similar songs sometimes.
I used to listen exclusively to Discover Weekly for a while. One week a finish rap song made it to the list, and I skipped past it every time it came on. Next week there was 2 finish rap songs. Then 5. Eventually half of my discovery list was finish rap, something I have no interest of. Canceled my subscription shortly after. Their algorithms are feeding themselves. I wish they had a dislike button so I could at least give them some sense of direction.
Did you know Spotify also has a its approve and disapprove buttons in the desktop app on opposite sides depending on what particular radio-like feature you're using?
I have the exact same thing with Dutch rap music. I'm not Dutch, don't speak Dutch and have never listened to Dutch music. But Spotify consistently puts 2-4 Dutch songs on my Discover Weekly.
I think this might be a huge problem for them and I’m pretty sure that they’re aware of it. The lack of variety seems to span all genres and radio stations. I’m really into Texas Country, Americana, and Red Dirt music. These are distinct sub-genres under “alt-Country” music but from the Spotify perspective, they’re all basically the same 40-50 songs. I can’t tell the difference between the Reckless Kelly, Robert Earl Keen, and James McMurtry radio stations because they all play the same set of songs that I got sick of months ago.
I don’t know if it’s a licensing issue—do they save money by keeping song variety down?—or is it an algorithm problem? I wish they would fix it because I’m about to bail for some other service if they can’t.
Pandora is the gold standard for good variety and new artist discovery as far as I’m concerned.
Pandora? Really? I've found Spotify to be orders of magnitude better than Pandora for both variety and discovery. Pandora was the first to have competent discovery functionality, and it was impressive for its time, but I'm wondering if this is "good old days" syndrome speaking.
With Pandora I can give it one song, and get a playlist out that won't repeat for a couple of hours (I really wish it was longer!)
With Spotify, I enter in a song, and get lots of stuff from that same artist (something Pandora can't do!) and then after that's done, a bunch of repeats.
Then again my problem may be that I'm trying to use Spotify the same way I use Pandora.
If you're looking for more noise to get you out of your music comfort zone then check out JQBX (https://www.jqbx.fm). It's a social music app that let's you DJ and listen to music with others in virtual rooms (similar to turntable). It plays all audio through Spotify so you get a huge library, can save tracks for later, and the random plays can seed your new discover weekly playlists (or you can use private mode so it won't).
FWIW I find probably 1/3 of my music through DW, 1/3 through manually browsing related stuff on Spotify (if I discover a new artist and see they were featured on a compilation, I might listen to that compilation) and 1/3 from real life (friends' recommendations, unknown bands from festivals etc). For my pattern of use, DW is varied and absolutely amazing. I only listen to my DW probably 1 week in 4 though.
I'm a former Last.FM user. They'd play a mix of past favorites and related recommendations, then adjust based on feedback (e.g. likes). There was a noise slider that would influence the ratio of new songs (and indirectly bring in recommendations from farther away from your core tastes). When I first found out I could be more adventurous, I was very excited, as I (like most people) pride myself on being curious. As it turns out, constantly discovering new songs from unusual genres is exhausting. This turned out to be a feature I very much wanted, but felt unhappy using and disappointed in me not using. Maybe this is common enough that the Spotify PM for Discover Weekly experimented and decided not to ship it.
I get constant metal versions of pop songs. Winterplay's Billie Jean was recommended to me for weeks on end. This week I have Turn down for what meets metal.
Seems like a small tweak to let us tell them we dislike this style of music. According to the article a -1 should be able to be put in that matrix (if that's actually how it works).
Also, I dislike that I lose music after a week. Which means I've lost a few good songs from their churn that I forgot to save.
It's nice the first few times, but after a while I get the impression that I'm trapped in a "Groundhog Day" loop and hear the same music over and over again.
Spotify should add a slider that lets me widen or narrow the 'search area', sometimes I want to hear more similar music, sometimes I want to find more stuff at the edges where all the interesting stuff lurks.
Spotify seems to mostly stick with the ~5 or so most popular song by a particular artist, so I am seeing some repetition. I don't mind, since I like most of those songs, but you should really take them as artist recommendations, and explore their other songs.
I've found, that if you put in more effort into discovering music yourself, Spotify's recommendations improve.
I used to browse the community forums, propose features and vote on others. Spotify, however, has been pretty unresponsive to even the most reasonable and popular proposals, sadly, I might say.
In case you didn't know, Spotify categorizes music into genres behind the scenes. You can use this site to find out what genres your favorite artists are categorized as: http://everynoise.com/engenremap.html
You can use this information to then check out Spotify's auto-generated playlists for each genre. They have at least three types for each one: "The Sound of <genre>", containing definitive representation of the genre, "The Pulse of <genre>", containing songs that fans of the genre listen to now, and "The Edge of <genre>", with unpopular songs (not necessarily of the same genre) that fans of the genre listen to.
This has been a great way for me to find new music that I like, especially "The Pulse". I even created a small script that takes a spotify playlist as input, parses all the artists, converts to genres, and creates a new playlist with "recommendations" based on the Pulse playlists, with each genre represented based on its percentage in the initial playlist.
Basically, I never use Discover Weekly, because I know it will eventually converge just like all my Pandora stations that cycle through the same 20 songs after a few months.
I only use the Discover Weekly list once in a while. I prefer using the Daily Mixes, and my own playlists. It seems to make Spotify match my taste quite well.
I've been using Discover Weekly for a few years now and haven't had that problem. The only annoyance that it sometimes adds a song I already have saved (and therefore have no need to 'discover').
> You can use this site to find out what genres your favorite artists are categorized as
That's my problem right there: my favorite song right now is a musician's piano cover of one of his own songs. His music is usually electronic, which I don't like, but I love this one song. So Spotify will recommend me electronic music from other artists, which of course does not fit my song.
Repeat for every author. I liked one song from a German musician, and now half the recommendations are German music. While I can see why network relations make sense, I wish I could say "give me a similar song, not a similar artist".
Usually on the "Sound of.." playlist there is a link to the Pulse and Edge in the description. Otherwise the genre might be too obscure and a playlist for it not generated. Also, the "Sound of" playlists are stored under the "Sounds of Spotify" account, while the pulse and edge under "Particle Detector", so you might try searching directly in that profile.
It's very impressive. But, I'm curious about some of the technical details. For each of these representations you basically end up with a dense vector on a song level. Which I assume you would then kNN with a user specific vector. But I've never come across a nice kNN data structure that supports high dimensional vectors in a larger than memory setting whilst supporting updates. Spotifys own Annoy is cool https://github.com/spotify/annoy, but changing or adding a song requires rebuilding the whole structure ... surely that's prohibitive at scale?
Spotify re-runs the latent vector models regularly and re-indexes them into Annoy indexes. There is no need to do that in real time, you can be a few weeks delayed and it's usually fine. New music doesn't have much data and need different methods anyway.
273 comments
[ 6.6 ms ] story [ 220 ms ] threadFind a popular song that is likely to be on a lot of people's playlists. Make a new song that is a close match to it for raw audio modelling. Launch on Spotify.
Yeah you'll probably only get one or two plays per user, but spread over a large number of users (who all listen to the whole Discover playlist every Monday, like I do) it's still significant traffic.
Or have I got this wrong?
Also, what you are describing sounds like what producers already do because people actually like it: find the latest trends in sound and copy them, (but hopefully with a fresh twist so that people get into it) :)
People who have music X on their playlist also have Y a lot. Person A listens to X but not Y. Let's make them discover Y.
You could just build a topology of songs like that and then recommend songs to user A if they are topologically close to the songs he likes.
EDIT: Read the article now :D they do that and it's called Collaborative Filtering.
(Not affiliate, just a big fan)
It's all gone down hill since they switched everything to mood-based radios. I feel like the main problem is that I like to listen to albums and Google seems to assume I only want to listen to random streams of disconnected singles. I seem to just get played the same stuff over and over, too.
I don't know how to find new (new to me, not the world) music anymore. Are there any good services that anyone recommends?
I have the same problem. Usually I find new stuff on the various music subreddits and last.fm's Similar Artists page. Even though I have tons of playlists and saved songs on Spotify, their recommendations usually suck (probably because I listen to just about every music genre under the sun.)
Their radio feature rehashes the same songs I've heard dozens of times, and Discover Weekly is usually so awful I never listen to it.
last.fm for me just recommended the next big artist in the same category. Often their music styles were still obviously different and the recommendation rather put me off. People that are hooked by the complexity, details and perfectionism in "Nightfall in Middle-Earth" won't necessarily like Manowar.
I recently got one Spotify recommendation that lead me into listening through a band's full catalogue and getting tickets for their show two weeks later (Insomnium, btw). It also dug out a song that I liked in primary school but completely forgot about. They discovered that I like cheezy metal covers of 80s pop songs and add one by some obscure band to my list every now and then. I'd say my experience was often pretty accurate.
In other words: yes, last.fm's recommendation was much too much based on "customers also liked", which helps in a lot of cases, but so often it horribly fails ("Customers who bought The Martian on Blu-Ray also bought this asthma medication because chance happens. Wanna try it out?").
I wish they kept it simpler with ratings, and explicit instead of implicit. The discovery weekly playlists are absolutely horrible for me to the point that I don't even bother checking them nowadays.
What works better for me for discovering new music with spotify is right clicking on an existing playlist and then "Create Similar Playlist" - that gives way more control over what kind of genre/style should the playlist consist of.
But I will go as far and say that implicit rating fits a lot more than explicit, because it doesn't require the user to do extra work on top of the base goal of listing to "good" music.
I feel like explicit rating scales very poorly with catalogue size. So when you have music, and as much music as Spotify has, then the work effort of explicitly rating your taste becomes too big, you start to not bother, and quality of rating becomes poor.
YouTube has good automatic playlists but the play-next feature often wanders way off after a few songs
It's big feature, among others, was their heavy rotation section - I only followed people that shared a similar taste to mine and we had a really cozy circle of listeners whose current favorites were surfaced by said heavy rotation section. We discussed albums in comment sections, shared playlists and I regularly stumbled upon familiar usernames and friends when discovering new gems. This social component to discovery is completely missing from Spotify, yet it's more powerful than any recommendation engine I have used, including Spotify's attempts, which I'd rate as mediocre.
I miss rdio.
One thing I worry about with it though is how much my behaviour might influence the choices. For example, if there's a track in there that I already know quite well, because I like it, I'm often scared of skipping it in case the algorithm takes that as a massive negative signal.
I believe that the concept of Frozen has been explained to her by her peer group, through the medium of hair plaits.
Three year old kids are a weird and amazing bunch.
Also, I have a young daughter who often asks me to play music in the car or Google Home and those affect my recommendations. I wish Spotify had some switch I could turn on to temporarily ignore anything I did until I switched it back off.
They've moved over to the Spotify API since I last had a play, but it's great that they still provide them.
You can get the audio breakdown of a track, as well as a summary of the track features including fun stuff like "danceability" and musical positiveness ("valence"). Radiohead's "Fitter, Happier" was low on both of these points if I remember correctly.
[1]: https://developer.spotify.com/web-api/get-audio-features/
[2]: https://developer.spotify.com/web-api/get-audio-analysis/
The neural network is trained to mimic collaborative filtering vectors from raw audio. It's a separate model from time signature, key, mode, tempo, loudness, etc.
I've tried training the algo by following artists and saving albums in the style that I would like, but these playlists keep peddling stuff that is way off the mark. Interestingly, the daily mix playlists have responded to this training, but not Discover Weekly or Release Radar.
For users who are tired of the same songs being fed to these playlists each week, you can create an IFTTT action to save the content of these playlists in separate archive playlists. Once a song is in the archive, it won't (or shouldn't) appear in either of the weekly so-called discovery playlists.
edit: grammar
They still do, sometimes, at least for me. But more importantly, adding songs to playlists sends the positive signal to Spotify’s recommendation engine, as far as I know. So, saving all your discover weekly tracks into archive playlists will encourage Spotify to give you similar music in the future.
Spotify’s “intelligence” in general is a huge let down though. Radio stations are extremely limited - more like 15-song static playlists indefinitely on repeat! It annoys me to no end, same for daily mixes. I end up listening to the same songs over and over and over and over again. Maybe that’s what makes them the most money?
Last.fm was amazing at finding me new music I liked. Rdio was amazing at.. radio :D I used to go for months on the same station. I miss both a lot, and occasionally I still use last.fm or everynoise.com to generate better playlists for Spotify.
I haven't noticed this bias at all. It might be due to our own music preferences.
But discover weekly only gives recomendations for the dominant genre.
For example at the moment Apple Music is suggesting to me four Wednesday playlist - all heave metal the genre I never liked and listened to. Also two artist spotlights - Jeff Chang and Danny Chan. Yes I have Japanesese account but this is Cantopop and Taiwanese crooner - so a little off the map especially that I do not enjoy pop music at all especially Asian. Some other examples are what I call comin denominators. Yes I follow lots of jazz but Frank Sinatra is not jazz music, nor Tony Benett. etc...
When I go to an filmfestival or an concert, part of what I pay goes to the proffessionals who make the selection for me. And I'm happy to pay their services, just like I pay for journalists.
Disliking multiple songs from the same artist only to have it suggest the next one from their repertoire is infuriating. Worse, Apple Music will play songs you've disliked in the radio anyway.
They have a ridiculous bias towards covers. I bet I've been recommended 50 (I wish that was an exaggeration) version of the Gladiator theme ('now we are free'). And they are all terribly bad (as in blood coming out of my ears bad). I am genuinely ashamed that someone thought that it would be a good idea to submit it to spotify - and if so how spotify could recommend it to anyone, there is no way spotify could have gotten any indication that anyone has ever liked any of those versions of that song. Similarly I get lots of game of throne covers, and now Despacito covers (never listened to that song in any variant on spotify, on purpose at least).
Turns out Spotify has a “Private session” mode to prevent influencing your taste.
Which would be quite bad, but that is how it often feels.
“Didn’t I” by Darondo seems to be a good example of this. A fantastic, forgotten soul song (and from the Bay Area!) that was assumed to be a recognition of my friend’s unique taste in his car... until the other four people inside revealed they’d all had it in the last two weeks, too.
It makes me wonder about the motivations. Did the IP for this song recently change hands?
(I still get the shivers recalling the month when people kept gleefully playing me “Temporary Secretary” by Paul McCartney; a new discovery from their Discover Playlists. I could have marched to Stockholm to strangle a data scientist.)
It could be popularized by DJs who are playing it, e.g. Four Tet, whose personal playlist on Spotify now has close to 39 thousand followers.
At least in that case specifically, I doubt it's due to distribution rights. There are no new releases indicated on Discogs, which seems likely if someone new acquired them.
It doesn’t surprise me that “Didn't I” comes up in many playlists, but it does surprise me that the other great songs on that album don’t get worked into Spotify stations. This is a great example of their algorithmic problems.
This is a really foreign concept to me. Why would you be let down because other people listen to the same song?
On the other it led me to double down on my purchased offline music collection, which I already maintained long before rdio.
- Go to the DW playlist
- Click the "..." options button
- Then to "Go to Playlist Radio"
- "Follow" the DW playlist radio
- You can now up/downvote songs
Recommender Systems for Self-Actualization https://dl.acm.org/citation.cfm?id=2959189
I don’t know if it’s a licensing issue—do they save money by keeping song variety down?—or is it an algorithm problem? I wish they would fix it because I’m about to bail for some other service if they can’t.
Pandora is the gold standard for good variety and new artist discovery as far as I’m concerned.
I'm pretty sure that'll shake up the recommendations a bit.
With Spotify, I enter in a song, and get lots of stuff from that same artist (something Pandora can't do!) and then after that's done, a bunch of repeats.
Then again my problem may be that I'm trying to use Spotify the same way I use Pandora.
Seems like a small tweak to let us tell them we dislike this style of music. According to the article a -1 should be able to be put in that matrix (if that's actually how it works).
Also, I dislike that I lose music after a week. Which means I've lost a few good songs from their churn that I forgot to save.
Do we know anything about the algorithms actually used?
https://www.google.fr/url?sa=t&rct=j&q=&esrc=s&source=web&cd...
My taste in music is also rather specific, which made it even more impressive. I shall see how the following weeks fare.
Spotify should add a slider that lets me widen or narrow the 'search area', sometimes I want to hear more similar music, sometimes I want to find more stuff at the edges where all the interesting stuff lurks.
I've found, that if you put in more effort into discovering music yourself, Spotify's recommendations improve.
I used to browse the community forums, propose features and vote on others. Spotify, however, has been pretty unresponsive to even the most reasonable and popular proposals, sadly, I might say.
You can use this information to then check out Spotify's auto-generated playlists for each genre. They have at least three types for each one: "The Sound of <genre>", containing definitive representation of the genre, "The Pulse of <genre>", containing songs that fans of the genre listen to now, and "The Edge of <genre>", with unpopular songs (not necessarily of the same genre) that fans of the genre listen to.
This has been a great way for me to find new music that I like, especially "The Pulse". I even created a small script that takes a spotify playlist as input, parses all the artists, converts to genres, and creates a new playlist with "recommendations" based on the Pulse playlists, with each genre represented based on its percentage in the initial playlist.
Basically, I never use Discover Weekly, because I know it will eventually converge just like all my Pandora stations that cycle through the same 20 songs after a few months.
That's my problem right there: my favorite song right now is a musician's piano cover of one of his own songs. His music is usually electronic, which I don't like, but I love this one song. So Spotify will recommend me electronic music from other artists, which of course does not fit my song.
Repeat for every author. I liked one song from a German musician, and now half the recommendations are German music. While I can see why network relations make sense, I wish I could say "give me a similar song, not a similar artist".
Proxies work.