4 ms·

My master‘s thesis was related to finding and implementing a music library vis tool, which should show relations between artists and songs by grouping them toge

by manmal 5y ago

My master‘s thesis was related to finding and implementing a music library vis tool, which should show relations between artists and songs by grouping them together. One important aspect I learned is that, in 2D space, there are only so many nodes you can add before the graph becomes useless. You could display nodes multiple times in the same graph, but this lowers usability a great deal. I‘m curious how this problem will be dealt with.

mattigames 5y ago
Zooming in and out seems to be their approach for that problem, the more you zoom out the most specific nodes lose their labels; for a music graph I would guess that would be the songs, and even further out maybe join artists by genre (or decade) But yeah that approach may be limited as well for such an ambitious project, in the long term I think they should have a list of categories you may check (or uncheck) to show only the subjects you are interested at the moment, e.g. health, money and education.
- informationart 5y ago
  Love this discussion and totally agree about the challenges of visualizing a network graph with lots of nodes! (I'm Lisa, head of design at System.) We'll be doing a lot more work on the graph in future releases to allow the user more filter controls.
  - EGreg 5y ago
    My dad got funding 24 years ago to make “InfoPlanet” and I think most of the time they worked a lot on visualizations of data as nodes on a planet. I also worked with a friend Alex 15 years ago who made TouchGraph and he had to work on that stuff too. I can put you in touch.
    - dr_dshiv 5y ago
      Alex is awesome!!
airstrike 5y ago
I remember what.cd seemed to have an amazing network graph for browsing through artists and discovering new stuff... sadly I think that's forever lost in RIAA-hell
- doublepg23 5y ago
  One of the great lossless in recent history. The current torchbearers don’t live up to wcd - I assume streaming is to blame.
- radlad 5y ago
  Redacted does something similar.
- greggsy 5y ago
  TBH I blame the advent of streaming platforms, and the public’s acceptance of brainlessly following recommendations from algorithms. I’m sure that each of the big platforms - YT, Apple Music, Spotify and Tidal - have someone in their team who wants to implement this type of thing, but they’re probably drowned out by overbearing algo marketing types. They need to take a long hard look at the popularity of the rock family tree publications, or at least consider the type of traffic that whosampled.com gets from people going into a relational rabbit hole.
trevyn 5y ago
The underlying problem is that the “real” graph of relations exists in a high-dimensional space; ML “embeddings” tend to tease out the actual structure of this space. Once you have a good high-dimensional representation, you can use something like t-SNE to reduce that to 2 (or 3) dimensions in a way that makes sense. But as you noted, if you use all the nodes, this dimensionality reduction loses quite a bit of information. The solution is to have a way of interactively limiting the number of nodes in an exploratory way, and having a dynamically-applied t-SNE running on only that subset of nodes. If you set the iterative parameters right, you can get some really fun interactive physics-like visualizations where nodes attract and repulse each other in intuitive ways. (“This thing is like this other thing in one dimension, but also not like it in another dimension” kind of stuff.)
- manmal 5y ago
  Yes, this is what I have been thinking - reducing high-dimensional data to lower dimensional is just super lossy. Funny that you mention physics-like visualizations - my implementation did use a springs (forces) model.
hyperpallium2 5y ago
Some kind of hierarchy - but nested abstractions seem difficult to name and unhelpful to use. What about picking a single instance (actual song) as representative? A platonic form, at that level of abstraction. Expanding that node would give "more like this" - consisting also of actual instances, representative of further sets.

by manmal 5y ago

mattigames 5y ago

Zooming in and out seems to be their approach for that problem, the more you zoom out the most specific nodes lose their labels; for a music graph I would guess that would be the songs, and even further out maybe join artists by genre (or decade) But yeah that approach may be limited as well for such an ambitious project, in the long term I think they should have a list of categories you may check (or uncheck) to show only the subjects you are interested at the moment, e.g. health, money and education.

informationart 5y ago
Love this discussion and totally agree about the challenges of visualizing a network graph with lots of nodes! (I'm Lisa, head of design at System.) We'll be doing a lot more work on the graph in future releases to allow the user more filter controls.
- EGreg 5y ago
  My dad got funding 24 years ago to make “InfoPlanet” and I think most of the time they worked a lot on visualizations of data as nodes on a planet. I also worked with a friend Alex 15 years ago who made TouchGraph and he had to work on that stuff too. I can put you in touch.
  - dr_dshiv 5y ago
    Alex is awesome!!

airstrike 5y ago

I remember what.cd seemed to have an amazing network graph for browsing through artists and discovering new stuff... sadly I think that's forever lost in RIAA-hell

doublepg23 5y ago
One of the great lossless in recent history. The current torchbearers don’t live up to wcd - I assume streaming is to blame.
radlad 5y ago
Redacted does something similar.
greggsy 5y ago
TBH I blame the advent of streaming platforms, and the public’s acceptance of brainlessly following recommendations from algorithms. I’m sure that each of the big platforms - YT, Apple Music, Spotify and Tidal - have someone in their team who wants to implement this type of thing, but they’re probably drowned out by overbearing algo marketing types. They need to take a long hard look at the popularity of the rock family tree publications, or at least consider the type of traffic that whosampled.com gets from people going into a relational rabbit hole.

trevyn 5y ago

The underlying problem is that the “real” graph of relations exists in a high-dimensional space; ML “embeddings” tend to tease out the actual structure of this space. Once you have a good high-dimensional representation, you can use something like t-SNE to reduce that to 2 (or 3) dimensions in a way that makes sense. But as you noted, if you use all the nodes, this dimensionality reduction loses quite a bit of information. The solution is to have a way of interactively limiting the number of nodes in an exploratory way, and having a dynamically-applied t-SNE running on only that subset of nodes. If you set the iterative parameters right, you can get some really fun interactive physics-like visualizations where nodes attract and repulse each other in intuitive ways. (“This thing is like this other thing in one dimension, but also not like it in another dimension” kind of stuff.)

manmal 5y ago
Yes, this is what I have been thinking - reducing high-dimensional data to lower dimensional is just super lossy. Funny that you mention physics-like visualizations - my implementation did use a springs (forces) model.

hyperpallium2 5y ago

Some kind of hierarchy - but nested abstractions seem difficult to name and unhelpful to use. What about picking a single instance (actual song) as representative? A platonic form, at that level of abstraction. Expanding that node would give "more like this" - consisting also of actual instances, representative of further sets.