The effect of representing networks with ‘spaghetti’ network graphs, like the above inscrutable graph I found on Google images, is surprising because simultaneously almost completely illegible and yet at the same time immediately satisfying. Whenever I show a network graph when describing my own work everyone seems to ‘get’ what I’m up to.

If you want to gather data for social network analysis, or check it, or edit it, you tend to do so using a matrix table. Doing so via a network graph is going be very hard.

If you want to boil your data down into some aggregate picture then you can use mathematical approaches to derive properties: modularity, connectedness, etc. If you try to guess these properties by looking at a network graph, your intuition is not going to be great.

And yet network graphs of my work seem to be incredibly important for people to be able to mentally situate what’s going on, to position what I’m doing in their minds. It seems to live in between the comprehensive tabular matrix and the reductionist statistical analysis and fill a unique, qualitative role.

Gephi, which I use to visualise networks as graphs, has various algorithms for creating the network layouts. They are computationally expensive and take several seconds to run, yet after all this computation the result often leaps off the page as visually wrong – unbalanced in someway. Usually I can see what needs to change to make it right.

The sense of orientation that comes from a seeing a network graph, and the immediate ability to layout a graph in a way that apparently a computer cannot might be linked at some cognitive level – do humans have a special module for processing network graphs?

In any case, what I previously thought off as a bug – the seductive quality of the spaghetti graph – I am now reconsidering as a feature – that network graphs, even borderline illegible ones, give us some kind of context and confidence in the data we are examining. Perhaps they just act as a handy prompt to ask some important questions: are there meaningful clusters? Is the graph complete? What are the nodes, actually? What types of edges are there?

Doing Better?

Some brief research has turned up some approaches to reducing the amount of spaghetti in the network diagram.

Screen Shot 2015-03-04 at 14.58.03

Above is an attempt to add matrices to network diagrams[1]. Representing both halves of a necessarily symmetric matrix violates all kinds of Tufte dictums, that aside this graphic fails because it doesn’t aid intuition very much at all, and the hard data is impossible to read off because the matrices don’t have labels.

Screen Shot 2015-03-04 at 15.16.42

In another approach [2] we see an attempt to make clusters more intuitive. I think this work is more successful than the first example because of the specific focus on a qualitative, at a glance approach, however it’s also representing fewer nodes and edges so perhaps has set itself a lower hurdle. In case you can’t tell, the various coloured clusters have been encouraged to form into recognisable shapes – square, circle, heart, light great is approximating a triangle. But why square, circle, triangle? Was the key problem with this diagram comprehending the clusters anyway?

 

Contextual, rather than algorithmic

In looking through data visualisation books I found pure networks, the kind that the two examples above are trying to represent, quite rare. But we’ve all seen a very famous example of network data vis – perhaps the canonical example of data visualisation – the London Tube Map. What makes it work so well is the judicious addition and removal of information. The Thames isn’t part of the network, but without it the stations are completely geographically unmoored. Yet the precise distances between the stations have been scrapped, a detail that gets in the way of the aesthetic.

The tubemap has been been done to death, so I’ve included British Airways’ network graph of their flights, circa 1989. Here there is an extra contextual detail of the dotted connections which reach around what would be the back of the globe if we were looking at normal map of the earth. I also like the pleasing way that different destinations peel of a central spine.

These bespoke visualisations seem to be pointing out the inadequacy of the purely algorithmic approach of software packages like Gephi.

Even so, it seems that we take something from even the worst spaghetti diagrams.

IMG_20150304_130106

 

[1] Henry, Nathalie, J. Fekete, and Michael J. McGuffin. “NodeTrix: a hybrid visualization of social networks.” Visualization and Computer Graphics, IEEE Transactions on 13.6 (2007): 1302-1309.

[2] Shannon, Ross, Aaron Quigley, and Paddy Nixon. “Graphemes: self-organizing shape-based clustered structures for network visualisations.” CHI’10 Extended Abstracts on Human Factors in Computing Systems. ACM, 2010.

 

 

Comments are closed.