Add FA2 option to stop after N iterations ?

[RouxRC]

It would be nice to be able to set FA2 to run for exactly a desired amount of iterations before autostopping, instead of having to use a setTimeout to stopForceAtlas2 which will result in various numbers of ran iterations depending on the client's machine and browser.

[Yomguithereal]

Rather than an option, the following method will probably be implemented:

s.runForceAtlas2Iterations(n);

[sheymann]

An alternative implementation is available in:
https://github.com/Linkurious/linkurious.js/tree/linkurious-version/plugins/sigma.layout.forceAtlas2

Setting autoStop = true will force the algorithm to stop automatically when a quasi-optimal layout is reached.

[paulgirard]

quasi-optimal layout being defined by :

       // Auto stop.
      // The greater the ratio nb nodes / nb edges,
      // the greater the number of iterations needed to converge.
      if (W.settings.autoStop) {
        W.converged = (
          W.iterations > W.maxIterations ||
          alldistance / W.nodesLength < W.avgDistanceThreshold
        );
      }

For people doing Visual Network Analysis, this autostop feature might represents a problematic bias.

[RouxRC]

Hi,
Looking at your code, this does not really fit my needs since it seems like a "magic" setting with no user-control on it. At least if maxIterations and avgDistanceThreshold were configurable settings it would be better, but still, I'd prefer as @Yomguithereal proposed a runNIterations function, since you might want to do that sequentially in batches, and not stop forever the spatialization algorithm.
Moreover, I'm not really confident using a fork that seems to have left the road so far away from the official code. I'd rather keep things robust and stay on the official maintained code.

[sheymann]

@RouxRC good point for the configurable settings, I'll add them.

@paulgirard The underlying problem is the inability to compute a "good number of iterations": too few the layout is not readable, too large you lose computation time. My approach is to stop the layout when nodes move too little, while keeping a tolerance to local minima. In a methodological point of view, the "stop" decision taken arbitrarily by a human is more problematic than having a public algorithm taking that decision, because such algorithm can be criticized, measured and improved, while the human decision is a black box called "expertise".

[paulgirard]

Well. Yes the problem is the inability to compute a "good number of iterations".
Mathieu Jacomy who wrote ForceAtlas2 tried to find a good way to define a methodology to stop iterations
He couldn't find any automatic/algorithmic way to define the moment to stop iterations that would work for any kind of network.
On complex networks this technique will not work.

Thus he chose not to add any autostop method in the paper about ForceAtlas published in PlosOne. You probably know that since you are a co-author.
We don't want to "sell" something that seems to be a methodological solution but is, in fact, just a side effect.

And yes, I do confirm that it's not a good idea to try to do that with ForceAtlas when you want to render any network.
And yes, I do think there is a methodological problem in doing that.
Letting the user decide when he thinks the algorithm reached an equilibrium or letting him decide how much time the algorithm need to spacialize at a certain level of quality has nothing to do with blackboxing.

Now, if you think that for Linkurious' clients, it's a good idea to do it, that's completely fine with me but some sigma.js users actually use it for science, in which case that 's highly insufficient.

Finally please don't try to bullshit me about blackboxing.

[Stale]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.