I don’t know if this is a known issue in the academic world of graphs because I am way in the periphery — I use only free graph visualization tools (a lot of good ones listed at the InfoViz wiki).
I’ve noticed that with 2D force-directed layouts, you get a lot of nodes spatially nearby each other which can be relatively far away in terms of edge distance. In this image, 1 and 9 are not connected but are placed next to each other – which is exacerbated at places like artandculture.com or Quintura where they hide the edges. Beyond hiding nodes based on locality, I was thinking a way to fix this would be loosen the edges or fade the nodes proportional to distance from whatever node is highlighted. Other ideas?
March 7, 2007 at 5:22 pm
There is a method where you lay out a graph in a high dimensional space and then project onto a 2-dimensional space using Principal Component Analysis. I haven’t tried it but I think it has the property that distances are very related to overall connectivity. And it wouldn’t get stuck in local optimum like the force based methods you normally use. Maybe Google will find an open source implementation of this.
The example you give I think most people would view as a feature. If two nodes are highly connected to lots of the same nodes, normally you would hope that the graph software would point out that the two nodes are in some sense close to each other by drawing them close to each other. Hmm… now that I think about it maybe the PCA methods would exacerbate this problem. If you do try them let me know how they go.
March 7, 2007 at 5:33 pm
Yeah – I guess the example is not so great. In my influence graph there are lots of people who appear in the center because they are being pulled on by one connection from one side and one connection from the other. So they should actually be on the periphery as they are distant from the central nodes. It would look a lot better if I would just pick one influence to be more important but I’m reluctant to remove too much information for better layout.