Oh, finally this thread is fine ^^ !
Participate here to the writing of a paper about ForceAtlas
Here's the deal:
- Everybody may participate
- I'm the developer of ForceAtlas, you are the users, so everybody has something to bring. You'll at least tell me what interests you to read.
- Everybody is enjoined to review the work in progress.
- Those that actually participate will have their name on the paper
- But there is no guarantee that it will work, or that the paper will actually be published...
Even if it would be really cool.
Here is our starting point:
Issues addressed in the ForceAtlas paper
- ForceAtlas is made to spatialize Small-World / Scale-free networks, meaning "real" data (and not theoretical lattices or geometric figures...)
- Performance and Quality: ForceAtlas' purpose is not performance, but quality. Some force directed layouts seek performance (Yifan Hu, OpenOrd) while other seek quality (LinLog, ForceAtlas). The ForceAtlas has been empirically developed to allow a rigorous interpretation of the graph (ex. in SNA), with the fewest biases possible, and a good readability, even if it is slow.
- The pattern of attraction / repulsion is similar to the "Früchtermann Rheingold" layout plus gravity (very classical), but...
- The "quality trick" that gives ForceAtlas its identity comes from its asymmetrical repulsion force, that takes in account the degree of nodes. Andreas Noack proposed the LinLog layout, that features an "edge directed" option that is very close to this principle. Noack formalized a quality measure that fits to the needs we had, and we empirically produces something similar to what Noack produced.
- But contrary to the LinLog, ForceAtlas has no strong mathematical foundation (it's the result of an empirical research). We issued stability problems due to the asymmetry of the force, that's why we developed the "auto-stabilize" function (that is different from the Barnes Hut optimization, and from the optimization in a layout like GEM, but there might be similar things in some layouts that I don't know).
- We claim that as far as the purpose of the layout is to interpret the graph, options that give different views are useful and that's why we propose the "distributed attraction" setting. Once activated, it divides the attraction of an edge by the outdegree of the source nodes. This makes hubs attract less while authorities attract more. This makes sense for some types of graphs like web networks, where we consider that authorities are more central than hubs (the "value" is the authority or the indegree). This features keeps authorities in the center and push hubs to the border.
- As well, we think it's useful to have a "nodes don't overlap" feature. This is not a mathematical feature, but a semiotic feature. Gephi is not a theoretical framework but a tool. These non-purely-mathematic feature are precisely what we want to get in a tool like Gephi.
Tell what you think about that !
Mathieu Jacomy - webatlas.fr - sciences po médialab