Tag: Gephi

Refining the Musician Networks, Part 1

I recently posted about the six new saxophonist networks I created using MusicBrainz data and Gephi, and have subsequently created another four, including networks for two of the acknowledged giants of the instrument – Charlie Parker and John Coltrane. However, as I was digging deeper into the data I realized that there are a lot of redundancies in the data due to a couple of grammatical issues. There are two major issues I have now addressed that will make for cleaner networks.

Here’s what these networks currently look like:

Snapshot of Dexter Gordon graph network

I expect to replace these with cleaner, more logical graphs after making this pair of changes. The end result will have fewer nodes and fewer edges crossing each other to connect nodes.

The first change to address is the subtle but important difference between an apostrophe character ( ‘ ) and the similar yet slightly different grave ( ` ) character. Each one is used to represent an apostrophe in the MusicBrainz release.name field, leading to duplicate entries that are actually the same song. For example, ‘Round Midnight versus `Round Midnight. Subtle difference, right? But one that my postgres queries and ultimately Gephi see as two unique songs, cluttering the network graph unnecessarily. So how do we fix this issue in the data?

I first created a new version of the releases table, just in case something went wrong as I tried to make any updates. We now have an empty table with all the same attributes as the original. Step 2 is to populate the new table with a simple SELECT INTO statement:

select * into public.release_new  from public.release r

The next step is a bit trickier since it involves an apostrophe character, which postgres treats as quotation marks for other characters. We have to use some additional formatting to convince postgres that we really do want to replace all our ` characters with ‘ characters. Here’s the code I used (there are several ways to do this):

UPDATE 
   public.release_new
SET 
   "name"  = REPLACE(name,'`',E'\'')
WHERE 
"name" > '0'

Without going into too much detail, we are telling our query to find all ` characters and replace them with an apostrophe. I recognize this might mess up some cases where there is an actual grave accent on a song name, but we will now have a consistent approach rather than two slightly different characters throwing us off. The goal is to ensure that we recognize a given song as a single node for our graph as much as humanly possible.

We have a second issue to correct for, but this one can be done within our query rather than updating the database. In this case, some songs are listed in upper case in one place, and then as lower case in another. We could force all text to lower case for a match, but that is less than ideal. The same holds true for upper case; we don’t want our graph labels to be all caps. A third solution is to use the INITCAP function in postgres, like this:

select b.id, b.name, b.label, b.type, SUM(b.size)as size
FROM
(SELECT DISTINCT INITCAP(a.id) as id, INITCAP(a.name) as name, INITCAP(a.label) as label, a.type, a.size

INITCAP forces all first letters to upper case while leaving the other letters as they were. It’s not a perfect solution; apostrophes cause us a problem here too, but it’s perhaps a 99% solution. By correcting the apostrophe format and then using INITCAP, we now have a much cleaner query result for Gephi. As an example, the nodes query for Joe Henderson now returns 269 records, versus the original 285, an improvement of > 5%. This should certainly help clean up our graphs, as it will also reduce the number of edges connecting the nodes.

In part 2 of this series I’ll show the impact these changes have on our network graphs. The beauty of these changes is that I can apply the logic to all future graphs. Thanks for reading, and see you soon.

Six New Saxophonist Networks

Decided to play with the latest version of Gephi by creating a new musician network, and wound up creating six using MusicBrainz data. This was a fun project and will be followed by additional work covering more great jazz musicians. Here’s a quick screenshot of one of the graphs showing the artist, album releases, and songs associated with those releases:

Dexter Gordon music network

I’ll post the links to each network below and then take a walk through the creation process. Note that the graphs are all interactive, with panning, zooming, edge removal, and other features all available. More on those features later in the post. Here are the links to each network graph:

Note that there are at least two major omissions among saxophonists – Charlie Parker and John Coltrane, and of course some other notable names. I’ll plan to address those omissions in a future post.

To create each of these graphs I followed a simple set of steps and then used the same settings to create graphs with a consistent look & feel. The goal is to have users focus on the structure and content of the network as opposed to having to deal with changing shapes, sizes, and colors. Perhaps I’ll alter this for different instruments – piano or trumpet, for example may have a different color palette. For now, the color palette I have used conveys an appropriately jazzy aura, with the dark background and contrasting pastel-like node colors and subtle gray edges connecting the nodes.

The data for this project is sourced from the impressive MusicBrainz database. Note that MusicBrainz data covers many genres beyond jazz, but for my current purposes the focus is on jazz. I have created a local version of the data using DBeaver for writing and running SQL queries to retrieve data for ingestion by Gephi. DBeaver is a great solution for me – all of my other databases are in MySQL, while the MusicBrainz data is in PostgreSQL format. No problem, as DBeaver can handle both types (as well as many other data formats) with ease.

Here’s an example of the code used for node creation for Sonny Rollins:

SELECT a.* 
FROM
((SELECT CONCAT(ac.name, ' (Artist)') AS id, ac.name AS name, ac.name AS label, 'Artist' AS type, 50 AS size

FROM public.artist_credit ac

WHERE ac.id = 21832)

UNION ALL

(SELECT CONCAT(r.name, ' (Release)') AS id, r.name AS name, r.name AS label, 'Release' AS type, COUNT(DISTINCT rl.release) AS size

FROM public.release r
INNER JOIN public.artist_credit ac
ON r.artist_credit = ac.id
INNER JOIN public.medium m 
ON r.id = m.release
INNER JOIN public.medium_format mf 
ON m.format = mf.id
INNER JOIN public.release_label rl
ON r.id = rl.release
INNER JOIN public.label l
ON rl.label = l.id

WHERE r.artist_credit = 21832 
GROUP BY r.name
)

UNION ALL 

(SELECT CONCAT(ta.name, ' (Song)') AS id, ta.name AS name, ta.name AS label, 'Song' AS type, COUNT(DISTINCT rl.release)
AS Size
FROM public.release r
INNER JOIN public.artist_credit ac
ON r.artist_credit = ac.id
INNER JOIN public.medium m 
ON r.id = m.release
INNER JOIN public.medium_format mf 
ON m.format = mf.id
INNER JOIN public.release_label rl
ON r.id = rl.release
INNER JOIN public.label l
ON rl.label = l.id
INNER JOIN public.track t
ON m.id = t.medium
INNER JOIN public.track_aggregate ta
ON t.name = ta.name

WHERE r.artist_credit = 21832 
GROUP BY ta.name
)) a

While the code may appear complex, it’s goal is simple – retrieve all releases and songs for the artist Sonny Rollins, who has the ‘21832’ id. This code creates nodes for the artist (first section), all releases (second section) and all songs (third section). It uses the UNION ALL statement to combine the three sections into a single output file.

We then run similar code to create an edges source file:

SELECT a.*
FROM
((SELECT CONCAT(ac.name, ' (Artist)') AS source, CONCAT(r.name, ' (Release)') AS Target, 'Artist' AS source_type, 'Release' AS target_type
FROM public.artist_credit ac
INNER JOIN public.release r
ON ac.id = r.artist_credit
INNER JOIN public.release_label rl
ON r.id = rl.release
WHERE ac.id = 21832)

UNION ALL

(SELECT CONCAT(r.name, ' (Release)') AS Source, CONCAT(ta.name, ' (Song)') AS Target, 'Release' AS source_type, 'Song' AS target_type
FROM public.release r
INNER JOIN public.release_label rl
ON r.id = rl.release
INNER JOIN public.medium m
ON r.id = m.release
INNER JOIN public.track t
ON m.id = t.medium
INNER JOIN public.track_aggregate ta
 ON t.name = ta.name
 
WHERE r.artist_credit = 21832)) a
GROUP BY a.source, a.target, a.source_type, a.target_type

This output will instruct Gephi to use the artist as a source node and all releases as target nodes (first section) and then to use all releases as source nodes with songs as target nodes. Think of this as a hierarchy of Artist –> Releases –> Songs where individual songs are associated with the release they appeared on. Of course, in jazz, many of the most popular songs will appear connected to multiple releases, ultimately making for a more interesting graph.

Now that we have created the source files, let’s shift to Gephi to see how we use them.

Gephi allows us to pull in spreadsheet files as long as they meet certain criteria. Node files should have a name, label, id, and preferably a size attribute, although this can be created within Gephi based on the data. Edge files must have source and target fields, and ideally a weight value corresponding to the strength of network connections.

Here’s our data after ingestion, starting with the nodes:

Dexter Gordon nodes in Gephi

I forget to mention the usefulness of having a ‘type’ column; this will make it simple to set node colors in Gephi. Now the edges file:

Dexter Gordon edges in Gephi

You can see the source and target values, which are critical to how the graph will be displayed. Our edge weights are all set to 1 in this network, but frequently we will have varying numbers to indicate stronger versus weaker connections.

Here’s our completed graph in Gephi, after using a number of settings:

  • Setting the node colors by type in the Partition tab
  • Sizing the nodes in the Ranking tab
  • Choosing a layout algorithm – Force Atlas 2 is a popular choice
  • Scaling the graph to an appropriate size
  • Preventing overlap of nodes

This process can be very iterative, playing with different settings until you are pleased with the results. For graphs like this with hundreds of nodes, different options can be tried very quickly.

Dexter Gordon graph in Gephi

The next step is to export the underlying data as a graph file – .gexf is my choice for the web template I use. Here’s a small subset of the Dexter Gordon .gexf file showing the name, type, and size associated with each node.

Data from the Dexter Gordon .gexf file

Next, we can update settings in the config.js file; These will adjust the display parameters for the nodes and edges; note that there is also a .css (Cascading Style Sheet) file where many more modifications can be made.

Set graph options in the config.js file

Finally, we have the index.html file that contains links to several scripts as well as the config file. This is where we can also add a title and small bits of information about the graph content.

The index.html file is used for titles, script references, and other display options

I’ll be creating additional network graphs using this same end to end approach. The process becomes easier once the code has been tested and validated, and the settings have been standardized in Gephi and the resulting output files; much of the effort will simply involve copying and pasting existing settings. Watch this space for new graphs, and thanks for reading!

 

The Music of Charles Mingus

In honor of the recently released Mingus live recording from 1973, Jazz in Detroit/Strata Concert Gallery/46 Selden, I have created a Mingus network graph detailing his many recordings and songs, using data from the MusicBrainz database. This new recording is not part of the network, but many other classic Mingus works are detailed.

This graph shows Mingus at the center, surrounded by his many recordings, with individual songs connected to the recordings where they appear. Songs are sized based on the number of recordings they appeared on, providing a quick glimpse into Mingus’ most notable tunes. Use the zoom and pan controls as well as the search box to navigate the graph. The data is sourced from the MusicBrainz database, while the graph is built using Gephi and sigma.js.

Here’s a link to the live version – enjoy!

ECM Label Network

One of the great strengths of the MusicBrainz database is that we can build networks using not only artists, but also record labels, or even individual songs. In this post, we’re going to explore an example (per a reader suggestion) using the ECM label and its variants. ECM is known for producing high quality recordings from an array of both jazz and classical artists. Keith Jarrett, Charles Lloyd, and Gary Burton are among some of the better known artists with multiple ECM recordings.

In contrast to our Miles Davis network, where the focus was on the artist, we now wish to see the labels (ECM) as hubs within the network. We’ll take a similar approach to constructing the node and edges files, although we now are going to create a multi-modal structure with 4 layers: Label –> Artist –> Release –> Songs. Using PostgreSQL, we can pull this data quite easily from the MusicBrainz database. Let’s start with the nodes logic:

SELECT a.*
FROM
((SELECT CONCAT(l.name, ‘ (Label)’) AS id, l.name AS name, l.name AS label, ‘Label’ AS type, 30 AS size
FROM label l
WHERE l.id IN(46800,1884,106711,123517))

UNION ALL

(SELECT CONCAT(ac.name, ‘ (Artist)’) AS id, ac.name AS name, ac.name AS label, ‘Artist’ AS type, 10 AS size

FROM public.release r
INNER JOIN public.artist_credit ac
ON r.artist_credit = ac.id
INNER JOIN public.medium m
ON r.id = m.release
INNER JOIN public.medium_format mf
ON m.format = mf.id
INNER JOIN public.release_label rl
ON r.id = rl.release
INNER JOIN public.label l
ON rl.label = l.id

WHERE l.id IN(46800,1884,106711,123517))

UNION ALL

(SELECT CONCAT(r.name, ‘ (Release)’) AS id, r.name AS name, r.name AS label, ‘Release’ AS type, COUNT(DISTINCT rl.release) AS size

FROM public.release r
INNER JOIN public.release_label rl
ON r.id = rl.release
INNER JOIN public.label l
ON rl.label = l.id

WHERE l.id IN(46800,1884,106711,123517)
GROUP BY r.name
)

UNION ALL

(SELECT CONCAT(ta.name, ‘ (Song)’) AS id, ta.name AS name, ta.name AS label, ‘Song’ AS type, COUNT(DISTINCT rl.release)
AS Size
FROM public.release r
INNER JOIN public.artist_credit ac
ON r.artist_credit = ac.id
INNER JOIN public.medium m
ON r.id = m.release
INNER JOIN public.medium_format mf
ON m.format = mf.id
INNER JOIN public.release_label rl
ON r.id = rl.release
INNER JOIN public.label l
ON rl.label = l.id
INNER JOIN public.track t
ON m.id = t.medium
INNER JOIN public.track_aggregate ta
ON t.name = ta.name

WHERE l.id IN(46800,1884,106711,123517)

GROUP BY ta.name
)) a

The four sections of code are united by one common attribute – the four record label identifiers associated with ECM. Section 1 creates nodes for the record labels, section 2 the recording artists, section 3 the releases, and section 4 the songs on each release. This should give us a very interesting network, although it will not have the same level of cross-pollination as the earlier Miles Davis network, as the songs are being associated with specific releases.

Creating the edges is quite similar, albeit requiring just three sections of code:

SELECT a.*
FROM
((SELECT CONCAT(l.name, ‘ (Label)’) AS source, CONCAT(ac.name, ‘ (Artist)’) AS Target, ‘Label’ AS source_type, ‘Artist’ AS target_type
FROM public.label l
INNER JOIN public.release_label rl
ON l.id = rl.label
INNER JOIN public.release r
ON rl.release = r.id
INNER JOIN public.artist_credit ac
ON r.artist_credit = ac.id

WHERE l.id IN(46800,1884,106711,123517))

UNION ALL

(SELECT CONCAT(ac.name, ‘ (Artist)’) AS Source, CONCAT(r.name, ‘ (Release)’) AS Target, ‘Artist’ AS source_type, ‘Release’ AS target_type
FROM public.release r
INNER JOIN public.release_label rl
ON r.id = rl.release
INNER JOIN public.artist_credit ac
ON r.artist_credit = ac.id
INNER JOIN public.label l
ON l.id = rl.label

WHERE l.id IN(46800,1884,106711,123517))

UNION ALL

(SELECT CONCAT(r.name, ‘ (Release)’) AS Source, CONCAT(ta.name, ‘ (Song)’) AS Target, ‘Release’ AS source_type, ‘Song’ AS target_type
FROM public.release r
INNER JOIN public.release_label rl
ON r.id = rl.release
INNER JOIN public.artist_credit ac
ON r.artist_credit = ac.id
INNER JOIN public.label l
ON l.id = rl.label
INNER JOIN public.medium m
ON r.id = m.release
INNER JOIN public.track t
ON m.id = t.medium
INNER JOIN public.track_aggregate ta
ON t.name = ta.name

WHERE l.id IN(46800,1884,106711,123517))
) a
GROUP BY a.source, a.target, a.source_type, a.target_type

Here we are simply connecting labels to artists, artists to releases, and releases to songs. Both the node and edge results are saved to .csv files for use in Gephi.

Once the data was in Gephi, I spent parts of a few days testing layouts, spacing, colors, sizing, and so on, before settling for the moment on using the popular Force Atlas 2 algorithm. I find it useful to start the process using the rapid (and less precise) OpenOrd algorithm whenever working with a fairly complex or large dataset. Then, once the basic network structure is revealed, we can move on to Yifan Hu, Force Atlas, or any of the more precise methods.

For the (for now) final version, I elected to size the nodes based on the number of outbound degrees, which will place more visual emphasis on the record labels and releases, respectively. Artists with multiple releases will also be represented by slightly larger nodes. So here are two versions, the first in Gephi:

The second version is from the web, after tweaking settings in sigma.js:

To interact with the network, click here. Thanks for reading!