Feedback on "Created for you" playlists

ListenBrainz
1

Hi!

I started scrobbling to Listentbrainz in 2023 (≅5000 listens that year), and developed the integration of “Created for you” playlists into Mopidy-Listenbrainz. (I posted about that in Support for recommendations in Mopidy-Listenbrainz).

Now that this integration has been merged to the master branch of Mopidy-Listenbrainz, I am tracking the “Created for you” playlists attached to my account.

First, I want to mention that it’s a great feature! Thank you for providing those playlists.

But I also want to share my doubts on the quality of the generated content: My music library has >600 albums, >9000 tracks with heterogeneous styles and origins, third of being French music. And I never see a French track in the “Created for you” playlists, nor a Jazz or classical track… Is there any documentation on the generation algorithm?

For last week jam, 50 tracks are from 27 (mostly mainstream, and great -I love them!-) artists: AC/DC, Amy Winehouse, Black Country, New Road, Black Sabbath, Bob Dylan, David Bowie, Iggy Pop, Johnny Cash, Joni Mitchell, Lana Del Rey, Led Zeppelin, Leonard Cohen, Nick Cave & the Bad Seeds, Nico, Nirvana, PJ Harvey, Pink Floyd, Róisín Murphy, Suicide, Talking Heads, Television, The Clash, The Doors, The Velvet Underground, Tom Waits, Wet Leg, Yo La Tengo.

But many French or Jazz artists belong to my favorite artists for the last week, month or year. It looks like the jam playlists never contain tracks I am listening too which aren’t popular worldwide. Is it true?

How to explain that? I am finally about to decide to stop listening to generated playlists due to their little varied content.

1 Like
2

Your lament has come up a few times before. As I understand it, the issue is not that the algorithm only suggests mainstream tracks, but rather that the user base is still small and therefore there are certain genres that have few people listening to them. This leaves the algorithm with little data to make suggestions for people listening to these genres. The solution is straightforward: advertise Listenbrainz to your friends and help increase the user base.

1 Like
3

Thank you for your reply and sorry for my poor English; my point was not to complain but to provide feedback and ask for details on the algorithm under the “createdfor” API.

If I right now randomly generate a one-hour playlist from my music library, I see the following artists: The Lounge Lizards, John Zorn, Pink Floyd, Plaza Franca, Jacques Douai, Einstürzende Neubauten, Tom Verlaine, Francis Poulenc, Nina Simone, Alain Bashung, Mademoiselle Berry, Alain Leprest, Colette Magny.

Whatever the number of friends I’ll convinced to join Listenbrainz, I doubt that it will result in a significant relative increase of Leprest, Magny, Douai or even Bashung visibility.

By the way, the documentation for the createdfor API doesn’t tell about the “user base” or any use of other users playing history.

That’s the reason why I was asking for documentation on the algorithm.

4

I am not a native speaker either, so I am very happy to learn you are in no position to see how dreadful my English is.

You are correct. This is something that has been explained by members of the LB team here at the forum (it was @rob if I remember correctly).

ETA: There you go: it was rob indeed. I also found the post of a user with the same problem.

1 Like
5

Thank you @biocv for the links! I’ll probably go back to my randomly generated playlists, and thus help diversify LB data.

1 Like
6

i think if the playlist would not only rely on other users plays but also add other bands from bandmembers of bands that you regular listen to, it would maybe spice up the data a bit and solve the problem that you have.

but this mostly requires an effort of adding the bandmembers of as many bands as possible.

also i don’t think you can have the suggestion stuff ONLY rely on listens from other people. it will recommand the songs that most people listen to, so they get listend to more and thus also get suggested more. and so everyone will get the same 100 songs suggested over and over. (because thats what we listen to most).

and having more listenbrainz listeners won’t solve that. because the obscure stuff will never stand out enough to get suggested.

2 Likes
7

Thanks everyone for chiming into this topic. I don’t have much new to add other than: Please encourage more people to use LB, and then we all get better data. :slight_smile:

4 Likes
8

i don’t think that that will be neccessary true.

it feels like the suggested for you playlist are only based on the most popular stuff.
more people will listen to popular stuff and thus the most popular songs will be suggested more and and therefor the popular songs will be listened more…
so the data will probably stay the same.

and in this way, smaller artists (that only have a few listeners) will probably never be suggested since they don’t have the following of a big listening audience.

it would be much nicer if smaller bands gets suggested much more over the BIG bands who don’t need the extra attention.

3 Likes
9

While there is some truth in this, it is very likely that an increase in number of users will improve recommendations. The Listenbrainz recommendation system is based on a collaborative filtering algorithm, which generates user-specific recommendations based on the listening habits of users with similar taste in music. Hence, if you loathe middle-of-the-road pop, you should actually receive few recommendations of pop music once these get purged by more suitable alternative suggestions. The problem described in the OP sounds like a typical statistical issue, caused by insufficient data.

10

i understand that the data of listeners is relevant and that more people (with a more boader music taste) could solve the issue.

the thing that i find annoying is that the bad recommendations isn’t tweaked to fix the issue when we have the current amount of listeners.

i think there are a lot of ways to recommend other bands (from data that is in music brainz)

  • suggest bands that have common bands members.
  • suggest bands that played a gig together (don’t use festivals, but if you have a main band with a support band, you van link those 2, especially if there is a tour and it is the same band + support for the whole tour)

this is data that we can get from music brainz that doesn’t require more listeners. and i think will give a better recommendation until there are enough listeners.
and i think it would give a better recommendation to @orontee who mainly listens to french jazz, i assume that the other bands of the members of the french jazz bands she likes to listen to, or other jazz bands they play with are more relevant that the mainstream music that gets recommended now

maybe it would be nice if there are multiple algoritmes that we can compare with each other.

but i get a bit frustrated that the only response is “once there will be more listeners it will work better”, until then it just doesn’t work, and i think new people who want to try listenbrainz for the first time will be like: “why would i use it, it doesn’t even work” and stop using it. and we will end up with the only people who use it now, and that are the die hard people who edit music brainz and hope the recommendations onne day will be better…

but as long as the recommendations thing isn’t fixed, i find it hard to persuade other people because they ask “what can it do?” and the only thing it currently can do is “recommend mainstream music based on your listening history”

3 Likes
11

There is more to ListenBrainz than generating recommendations, like the visualizations of listening history and the other exploration features, so I think you are being a bit unfair in your characterization of LB. Also, in my experience, the LB team listens to feedback and is willing to consider user suggestions. So it’s good that you share some of your ideas here.

That being said, I think the strength of the recommendation system is that it comes up with suggestions that do not follow from obvious links like shared band members. Personally, I treat the exploration lists as a complement to other means of discovery and am prepared to accept that not all of its suggestions are that great. And yes, I did manage to discover some new artists that way.

Finally, I don’t think everybody will experience the problem as severely as the OP. My recommendations aren’t all that bad. Surely, they contain the more established representatives from my corner of the musical universe (Aphex Twin, Massive Attack) but it is clearly geared towards my musical preferences. My impression is that most users will manage to get something useful out of those suggestions.

3 Likes
12

What other exploration feature? I think there is the for you playlists and then there is neighbourhood. And i think both of them use the same suggestive data.

but don’t get me wrong, i do appreciate the work that gets deliverd.

i also never say that we should stop using this recommendation method.
but i do think that adding different ways of linking bands together other then ONLY via user listens could benefit recommendations.

that why i suggest different way band can be linked. and with more different method, different people will have different ways to discover new exiting bands.

13

100% this – thanks for the giving a succinct synopsis of the problem/challenge.

In fact, when we notice that data skews towards more popular artists, we work hard to try and normalize the data to try and remove that bias. Last.fm orginally coined this as the “The Beatles” problem for when calculating similar artists data. Everyone listens to some Beatles tracks, but that doesn’t mean that Tupac is similar to the Beatles. We have more work to do on this front, but it does not affect the “created for you” recommendations, this problem is more likely to surface in LB Radio.

1 Like
14

That is not correct. Weekly jams/exportation use collaborative filtering data and music neighborhood uses similar artist data. It is true that both are based on the listens, but they are two distinct data sets.

We also have LB Radio and Hue Sound for discovery features.

15

So does the similar artist data not use the artist tags? I was hoping that I would improve recommendations by adding tags.

16

Similar artist data is not derived from artist tags, no.

However, adding tags (for artists or any entity, really!) is a good move, because tags will increasingly be used by other features. Right now adding tags will make the LB Radio feature better!

3 Likes