Analysing the dump has been finished and all of the reverted/reverting edits (metadata only) is a 2.7GB file. I need to download it from stat1005 and run some scripts on top of that to make some histograms and data.

This is result of the analysis (note that I omit any revert that took more than 48 hours in average):

I think we need a plot of this data.

I'd also suggested using the geometric mean for looking at time-to-revert. E.g. geometric_mean = function(x){ exp(mean(log(x))) }

In python, I'd do:

>>> from statistics import mean
>>> from math import log, exp
>>> 
>>> def geo_mean(x):
...     return exp(mean(log(x_val) for x_val in x))
... 
>>> 
>>> mean([1,2,3,4,5,6])
3.5
>>> geo_mean([1,2,3,4,5,6])
2.993795165523909
>>> mean([1, 100, 2, 3, 76, 88])
45
>>> geo_mean([1, 100, 2, 3, 76, 88])
12.605921135923992

This is the new data:

Another data:

(Same structure but for users who have reverted more than five edits per month)

We discussed getting the plots cleaned up and adding English/Spanish Wikipedia at our sync meeting.

This is great. Please graph the results, write a report, and give a description of the weirdness in Spanish's dump files.

This is the plot of geographical mean of revert time in Wikidata:

Make a report! Start from https://meta.wikimedia.org/wiki/Research:New_project :)

Started this: https://meta.wikimedia.org/wiki/Research:Revert_time_analysis
Will work on it more in the next couple of days.

Ready to review

OK I made an epic to cover other work we should do before we speak publicly about what we found. See T200898: Analyze the effects of ORES deployments on counter-vandalism behavior

I think this will make for a great follow-up paper to When the Levee Breaks. But more immediately, we'll get a better view of what counter-vandalism looks like on various wikis.

Oh! And to the point of reviewing this specific task, please limit your aggregate analysis to 12 months. This will help account for seasonality. E.g., December/January and September look weird and can appear twice in a 17 month sample.