Join GitHub today
GitHub is home to over 28 million developers working together to host and review code, manage projects, and build software together.Sign up
Partition percentages incorrectly composed across filters #637
First, this is a critical bug: Gephi is giving incorrect results without warning, undermining confidence in any of its computations with filters.
This is a little hard to explain, but the gist is that if you (1) turn on a filter, (2) compute some percentages, and (3) either change the filter or turn the filter off, change it, and turn it on again, then computing the percentages of step (2) under the new filter will give drastically incorrect (too small) results in a manner that suggests that the % under the second filter is being taken from the results of the first filter, not the global total.
That is, certain computations behave as if the filters are composed rather than changed.
Specifically I see this with the following sequence (there may be others):
I have an Attributes/Partition/Modularity Class filter defined.
A1. I select the first modularity class as the filter (display only that class).
I then repeat at #1:
Here of course I expect the percentages in B3 to be of the class chosen in B1, i.e., they add up to 100%. But they do not: they add up to 15.45%. Gephi seems to be taking a percentage of an already reduced set.
I see this as a CRITICAL BUG that should be fixed immediately, because it means Gephi is giving incorrect results without warning, greatly UNDERMINING CONFIDENCE in anything it does.
ONLY the first percentages computed are correct. The only way I can avoid this bug is to CLOSE THE PROJECT, read it in, and start over for EACH of my over 200 partitions. This is a large graph that takes about a minute to read in and display.
As an additional note, I wish I did not have to do steps 1-3 manually for over 200 partitions. Isn't there a way to output how node attributes are distributed a cross a class? I was able to Group the partitions and run various stats on this meta-graph, which is nice, but I can't figure out how to get it to run descriptive distributional stats on attributes I have defined (i.e., node type).