#metabrainz

no, scaling is of help but its hard to quantify how. i have two other observations we can discuss then

scaling, not full normalization?

ruaok: i only tried scaling till now, as we discussed the other day

do you think we'll get an improvement going further?

1) i got scores like 0.4 and 0.2 so on on a scale of 0 to 1 when i used a smaller set of listens

2) these scores are somewhat in line with using the ones we got when we were using counts directly

hard to say ruaok, the question is how to tell if the recs became better

_lucifer: what is mean is say I have two rec A and B, before scaling if the scores were 1, 2 and after scaling they are 10 and 20, does this mean any improvement?

pristine___: the scores are relative and not an indication of whether the user will like it, so a strong maybe

the scores make sense relatively but absolutely they are just a bunch of number

I understand they are relative, just trying to understand if we really need to scale/normalize

:)

pristine___: imo we probably should because that theoretically brings all users on a same scale but practically to decide that we need some way to see if tell whether the recs have become better or not

48 slides. ha!

[musicbrainz-server] reosarevok opened pull request #1725 (master…MBS-11141): MBS-11141: Convert Edit Relationship edit to React https://github.com/metabrainz/musicbrainz-serve...

MBS-11141: Convert Edit Relationship edit to React https://tickets.metabrainz.org/browse/MBS-11141

pristine___: what is the usual rmse for the model ?

I don't remember :( You can generate model HTML file for that it had a table

Has*

i did, its 0.12

i wanted to know the value without scaling

I get that, but I don't remember:( I will search for the html in my pc once I reach hotel :)

np, thanks

_lucifer: any talks you preparing

no, shivam-kapila , you ?

Wanted to, but couldnt find anything interesting

> imo we probably should because that theoretically brings all users on a same scale but practically to decide that we need some way to see if tell whether the recs have become better or not

Yeah _lucifer. Even if they theoretically bring all users to same scale, I think we should only work on this further if recs are actually improved for that we need a way to determine if recs have improved or not.

Improved recs is rather a user dependent term

Feedback is the best here

shivam-kapila: of course! Didn't pay attention to the best term here but I guess _lucifer understood 😉

yup :)

pristine___: just suggesting

🌹

[musicbrainz-server] mwiencek merged pull request #1724 (beta…mbs-11140): MBS-11140: Build Jed language bundles in prod https://github.com/metabrainz/musicbrainz-serve...

MBS-11140: Client side language bundles are not built in production https://tickets.metabrainz.org/browse/MBS-11140

shivam-kapila: Hi! About the folow button, I think it looks nice. If you want to try to hide text until the user hovers (you'll need to add icons there surely), check out the existing code: https://github.com/metabrainz/listenbrainz-serv...

I filled Saturday’s timetable with “State of stuff” topics, Sunday can be for more sessions but there are only MB ones for now, more can be added to the wiki.

Mr_Monkey: hi

updating beta.mb.o

I wanted to know what you prefer

Those icons ones or only text ones

I thikn riker would dsilike that

(or possibly *like* who knows :3)

I was confused as new ux concepts say to put less transitions

So that it doesnt steal focus

is anyone into pictionary? or only sticking to among us

Scribbl is fun

yea no ads as well i think

maybe close people can play pysch lol

Havent played among us

shivam-kapila: Ah, I see. Well, personally I like the transitions, and that the button is small when not hovered. But you could also just resize your follow button and have a similar result without the need for transitions.

Lets stick to what iliekcomputers suggested

Mr_Monkey: was it you who brought the projector last year?

Thanks for feedback

ruaok: Yis

planning on bringing it again?

Oh, shivam-kapila, one more thing: the button doesn't look vertically centered with the user name, as far as I can tell from a quick look

ruaok: sure, if we need it no problem

I think it might be nice to have, if you dont mind.

Mr_Monkey: old one or proposed one

The screen you shared yesterday (was it yesterday?)

Okay

ruaok: Sure thing.

thx!

This guy shivam-kapila : https://usercontent.irccloud-cdn.com/file/81vM8...

Gotcha

iliekcomputers suggested not to change the button near username

But only in feed list

Ah, I see

ruaok: do you want review feedback on troi too? I feel the answer to most of my questions of the form "why did you do it like this?" will be "because that's how it organically grew"

I can contribute changes in, but I'm not sure how much of it is in flux

yes, I'd like that feedback. there are a lot of things still in flux, but I think things are starting to stabilize.

I forgot that I had already added troi.acousticbrainz.annoy -- so that part is already done.

yep, I just have a few parameters to change there

I'm writing a manage.py-like entry script as we speak.

and I'm just putting in a new distance metric now

I remember that we had this discussion about BPM. for reasons, we currently treat integer multiples of a bpm as the same (e.g. I just tried with a 71bpm track and got back a 142bpm track as very similar)

peak finding in frequency space?

yeah, basically

people aren't really good at distinguishing them in many situations either

I wonder how many real world scenarios that will go wrong in.

I look forward to discovering that bug.

I'm trying to remember where I encountered this, I think someone I know was building a dataset of known tempos in order to evaluate an algorithm

get many people to annotate each song, etc, in order to be sure that the annotation is good

and some songs there was a 50% split between people who thought it was 60 and who thought it was 120

I'm trying to look it up

and these weird mutliples started showing up...

anyway, I think that when asking for songs of bpm 120, people would rather have a song _not_ show up (because the algorithm incorrectly chose 60 instead of 120) instead of having a bunch of obviously 60bpm tracks show up as "close"

agreed.

ah, some interesting feedback from MTG

> you should take other variables into account then. tempo is highly correlated with genre

> don’t recommend techno 60BPM to people who listen to ambient with 60BPM

sounds like our hybrid similarities (that include more than 1 feature) are going to be necessary soon

like building a content profile ?

alastairp: that's totally duh in hindsight. :)

yeah, but you can see how starting with an observation (tempo octave errors happen), and going down a particular path gets you in trouble

and you forget to look at the broader picture

https://www.researchgate.net/publication/343268... this is recent work by a researcher who I know

I would stay that there are many many paths in recommendation work... most of them get you into trouble.

(it's 200 pages, you don't need to read it all)

in section 6.6.3 he proposes a new metric for comparing tempo estimation error by using log2 scale, so that the scale of the error is the same in different octaves

most people evaluate algorithms using "if it's within 4%, it's good". but 4% of 60 is not the same as 4% of 120!

ding

I'm getting some good suggestions in another chat, which I'll note down

it also depends on if we want to do "similarity", or just plain search

I wonder if it makes sense for these features with fewer dimensions to use annoy, or if we actually just need solr

BPM and key?

yeah. "get me things with feature x" should just be search

I think they are a poor fit for annoy. the results end up in too few buckets.

+1

when you need to introduce fuzzyness into it, that's where it excels

yep, with range data.

diru1100: I'd love some other game than fightinggame tbh

"why did you do it like this?" will be "because that's how it organically grew"

this needs to be quoted for truth. mb 20 yr history, that is the answer to a lot of newer users question :D

diru1100: https://skribbl.io/#collapseHow does look liek fun! reosarevok wanna play this isntead/inaddition to this among us thing?

It's pictionary

You draw the word that you got and others will guess