fresh releases: the only thing that actually seems to function.
i understand that not everything is fully functional yet. and that there are plenty of ideas (good ideas).
but its a bit frustrating that the features that are already broken for months are being presented as the solution for the spotify thingy.
if everything would we working this would be a valid post. but since almost everything on the list is broken/not fully working i think it was to early to post this blog post.
first have everything working, then brag about it. don’t premature brag about stuff that doesn’t function.
if you don’t know about what blog post i’m talking, it is this one:
As has been explained to you previously, certain features that you lament about in your post (in particular the recommendation system) will automatically improve as the user base grows. Hence, luring more potential LB users in by advertising the available features seems like a very timely thing to do.
Similar artists, seems like a bug. Check for tickets?
Recommendations for users: Maybe you are more mainstream than you think. I keep myself to electro dubstep and it is still non mainstream stuff. Of course I’d love to see smaller artists getting recommended, but that’s probably another feature entirely
Artist generation: see next point
Popularity data: If there’s anything to blame here, it’s the frozen global listen counts. No listen counts, not top recordings. Not top recordings, not recommendation data.
I do aggree that this is a big issue that should be prioritised more, as you can see the global counts almost everywhere.
and i understand that the current algorithm needs more data to fully function. but that sounds like an excuse to not provide a better solution for current users.
can we experiment with a different algorithm?
but the main thing about this is not that the algorithm doesn’t work.
the main thing is that there are structurly things not working. but the blog postes about this like there is no issue.
In my experience, the LB team is open to suggestions, so if you wish to see a hybrid recommendation system implemented you can always request this feature. I did not see you make this request in your OP though.
TBH, I am not sure what is is that you are trying to accomplish. You seem dissatisfied with the state of certain LB features and you seem to be aware that a growing user base will remedy some of those problems. But instead of trying to get more people involved in LB you tell off the LB crew when they do make an effort to attract more users. Don’t you think you are undermining your own interests here?
im still not conviced that having more people will result in better suggestions.
the only way more people will result in better suggestions is if those people ONLY listen to specific genres en thus create more relevant data.
but if just more people submit the same listens you will end up with just more of the same data, resulting in more of the same suggestions.
the thing i want to get to is that there is structrurly stuff broken. and al those fancy features that are being used to lore new users in don’t function as expected.
the first impression is the most important one. and if new users have a bad first impression, that will stay with them.
so if you want to get new users, first fix the stuff that all the other features rely on, before promoting it as the big solution to all your spotify frustrations.
and this is something that i had said before, if the there currenlty isn’t enough data to give better suggestions than just mainstream music. maybe there should be a different algorithm that gives better suggestions on smaller data sets.
i would love to be able to brag about how good listen brainz is. but to be honest it currently isn’t. because the global track count is frozen, the popular artist tracks don’t work. because the artist popular tracks don’t work the similar artists doesn’t work. because the similar artist don’t work the playlist generator doesn’t work.
so the only selling point is that you can see your current listening history (which have to keep a close eye on since it sometimes matches an incorrect song)
so tell met what is the selling point of listenbrainz at the current state, and do you want that to be the first impression for new users?
Ah, I see. In this you are up against expert consensus I am afraid. Collaborative filtering approaches are widely used and well studied and they are known to suffer from lack of data. I believe your point about user diversity is valid though.
As I said, it is fine to request hybrid recommendation systems. I suspect these are very labour-intensive to develop though, so don’t get your hopes up. The easy solution is attracting more users and if you discourage the team from attracting more users until the recommendations improve, then you are trapping them into a Catch-22 .
LB has its issues and hiccups for sure and there is no harm in giving feedback to the team. In fact, that is usually warmly encouraged. But perhaps you can be a bit more constructive and suggest actual solutions for the issues you mentioned?
i understand that there are hickups but stuff that is so critical for the functionallity of all the other features should not be broken for so long.
and i would like to appologise if my language was to harsh. i know that i am not the best with words sometimes (mostly when i am frustrated).
so what is the nicest way to ask if they can please fix the global listen count so the popular track functions again AND the similar artist functions again AND the playlist generator will function again.
I somewhat agree with this opinion -not necessarily agreeing with LB being “bad”-, because the blog post gives a too generous image of the current state of ListenBrainz.
AcousticBrainz was shutdown years ago because it was supposed that Collaborative filtering would be the end to all our problems. It has not happened and nowadays audio features are the thing needed for any genre/similarity analysis or AI processing…
Spotify has closed their API to retrieve audio features, so now there is no available source for that… AcousticBrainz was in fact the alternative to Spotify in that sense, but since it was shutdown… this was clearly a really bad move.
There were critical voices years ago pointing to the situation we are living today. If not enough features are working or they are not interesting enough for developers, they are not going to be integrated out there… and therefore there will be not enough users using ListenBrainz.
For ex. you can connect ListenBrainz to Spotify, but exporting playlist to Spotify doesn’t work properly:
If every-time I create a playlist, it’s missing 20 tracks… then I spend more time adding them by hand than directly creating the playlist there. This is just an example.
Therefore… If the idea is to get more people:
We need enough reasons to recommend ListenBrainz to people using already Spotify/Apple Music. What will they get here not available elsewhere? Is there something better? If we can not positively answer this question, then clearly something is misaligned with the aim of the project.
Is ListenBrainz offering something different to devs than current closed source data-based APIs? It was highlighted that Spotify has closed its api, but does LB really offer an alternative?
Collaborative filtering is clearly not filling the gap right now. Is there something else considered?
Back to topic, no matter if we like the tone used by some people or how criticism is written, there are some points on this thread which should be addressed. And unless those 3 questions are answered in a positive way and some past decisions are reconsidered, the project will not grow.
We never claimed that. The data in AB is rubbish and there is no point in maintaining a trash heap of data that can’t be used reliably in any way. LB is not a replacement for AB, period.
@sanojjonas
First of all, thank you for the feedback. It may have sounded harsh and people might not like it, but I think in essence there are very legitimate questions in your posts.
I believe it boils down to a couple of main issues:
broken/unstable features
not enough data, especially for smaller artists
I’ll address the second issue first: for the most part, there’s not a lot we can do except wait for the ListenBrainz user base to grow, and with it the number of listens and variety of their taste.
We do however have plans to merge in some of the richer metadata we have in MusicBrainz and see how we can use that to reinforce recommendations, with the hope that we could recommend less popular artists.
Spotify has ~600 million active users monthly. We have 34 thousands, so the data we have will inevitably be much more sparse.
Taking your example of WIXCAKES, with a whole 3 users having listened to them we can’t do much in terms of recommendations. There’s just not enough data for it to be used. Recommendation algorithms are very complex and there’s no magic wand we can wave to make it work better. They need statistically significant data to work, and if we don’t have that we can’t recommend an artist.
As for the stability of the various features of LB, that is very much at the forefront of our preoccupations and we are working as we speak to improve the reliability of those services.
It has been frustrating for us devs too, playing whack-a-mole and taking time to fix services every other week, so believe me when I say this will be improved.
Do keep in mind that we are a very small team of 4 devs working on LB development (while also working on other projects), and for some features only a single person fully knows the ins and outs enough to be able to fix some hard bugs, or to rewrite whole features.
Fixing our bigger infrastructure issues takes time, and we deploy our progress as we go.
While commercial ventures like spotify will work on a new feature with 100 developers for a year behind closed doors and hatch a fully fledged product, we take the road of slowly improving all the features over time.
Does that mean our APIs are less powerful than Spotify’s?
Yes; how couldn’t they, when they sink more than 1 million euros a year in research VS. our team of 4?
But the LB APIs are here for foiled developers who now have a dead project on their hands, and we certainly won’t be pulling the kind of dick move that spotify just pulled on developers.
We aim to work with developers rather than focus on a monetary baseline for some investors.
I will say that probably part of the issue with recommendations is Spotify definitely does stuff with genres in the recommendation department that ListenBrainz doesn’t do at all. I listen to a ton of country and my Spotify recommendations reflect that. My ListenBrainz weekly recommendations has just four country songs in it (all very popular artists) but dozens of songs from popular artists (Britney Spears, Linkin Park, Ariana Grande, etc.).
I do agree with op though that while ListenBrainz is a cool project and will get better, it’s not at all close to a replacement for what those developers lost access to.
Tbh I don’t even care about recommendations much at this point, I’d just like ListenBrainz to fix the basic features it has currently. Focus on the basic features before trying to get into recommendations.
According to ListenBrainz stats page Brat was the most listened to album this year. But like mentioned above, if you go to Brat’s page the individual listens are missing for most of the tracks. Apparently this just got fixed while I was writing this comment thanks to your post lol. We’ve been reporting this for months now, so glad to see it’s no longer an issue.
The bug still exists that caused several of Brat’s songs to map to a completely wrong recording, putting that recording at the top of the year’s most popular recordings. Since it’s so prominent, would be nice to get those bad maps manually fixed if another solution is too hard rn.
Playlist exports to Spotify break if the descriptions are too long (which makes all ListenBrainz generated playlists except the daily break). Just a basic truncate would be a quick fix for that issue.
edit: Can someone check if Weekly Jams and Weekly Exploration export works? I can see that “Last Week’s Exploration” and a few other playlists still have descriptions of 300+ characters. If the issue has been solved for the weekly playlists then I will shorten the other descriptions
Tried both, both are still broken. You are correct that the length problem is fixed for those two, I think this is a second error. This time it’s happening because Spotify does not support new lines in descriptions. A new line anywhere in the description will break the export, even if it’s at the end with nothing after the new line. So add a find and replace for new line characters to the truncate I guess lol.
Also, as a side note, the HTML tags in those descriptions get copied over exactly as you see them in the editor. (screenshot is of a playlist in Spotify that I got to work by removing all new lines from description)
.
