4

I'm trying to use python to output a PNG of a force graph using networkx and matplotlib. Here's what it currently looks like:

networkx/matplotlib force graph

The nodes overlap quite a lot, the edge length is very short, and the clusters are very sparse.

Here's what happens if I visualise the same graph using d3.js (ignore the different colours, edge weights, and arrows, which I'm not trying to reproduce):

d3.js force graph

As you can see, this is much more 'legible' and visually appealing, because the nodes are spaced out better, and the clusters are closer together, leaving less white space.

Despite reading every topic I could find here and trying several dozen iterations with node size, figure size, edge weight, and an edge length property (that doesn't seem to do anything, but that might be because it's an arbitrary property), I can't make the matplotlib graph look anything like the d3.js one, even though both graphs were created with exactly the same nodes, node sizes, and edges.

Unfortunately, creating them all in d3.js isn't really an option, as I'm hoping to create tens of thousands of static PNGs, and I don't really fancy rendering them all in my browser then screenshotting them!

So, my question is: is there (a) a way to improve the matplotlib layout (by making edges longer, and/or avoiding node overlap, and/or decreasing whitespace), (b) a way to output a d3.js force graph to PNG (or any static format), or (c) another library I could use to output better formatted force graphs as images?

Chris
  • 151
  • 10
  • Update: I have tried the `iterations` parameter, which shows what I presume are different points in the animation of the graph, e.g.: `pos = nx.spring_layout(G, iterations=100) nx.draw(G, pos)` This _might_ be useful, but I'm not totally sure how I'd work out what number of iterations is ideal for different sized graphs... and, again, d3.js does this all automatically. – Chris Jan 14 '18 at 21:59
  • 1
    Also, I'd love to understand why someone downvoted this post, given it very clearly outlines the problem and the question! – Chris Jan 14 '18 at 22:01
  • Probably related to: https://stackoverflow.com/a/53156709/2912349 – Paul Brodersen Dec 21 '22 at 13:20

1 Answers1

3

I've found a way to make this work, by using graphviz_layout() rather than spring_layout() (thanks to a tip found here). This requires installing some extra libraries, so I'll document what I did (on Ubuntu 16.04) in case anyone else is having a similar issue.

To use graphviz_layout(), you have to import it first (from networkx.drawing.nx_agraph import graphviz_layout). However, you also need to install pygraphviz, which seemingly isn't installed as a dependency.

To get pygraphviz installed in a way that NetworkX could see it, I needed to do as follows (thanks to a post on this thread):

sudo pip install pygraphviz --install-option="--include-path=/usr/include/graphviz" --install-option="--library-path=/usr/lib/graphviz/"

Visualisations with graphviz_layout() seem to have less node overlap than spring_layout(), as well as much more æsthetically pleasing spacing in general. See below for output (compare to the first image in the question post):

working networkx/matplotlib/pygraphviz force graph

This took a lot of fiddling around to get right, so I hope it's helpful for others with similar requirements!

Chris
  • 151
  • 10