Global Security and Intelligence Studies Volume 5, Number 1, Spring / Summer 2020 | Page 63

Discovering Influence Operations on Twitch.tv: A Preliminary Coding Framework
speed of each stream was then
calculated.
a. Mean comment speed was defined
as the average of the users’
comment speeds calculated
earlier (only users with at least
three comments).
b. Median comment speed was defined
as the middlemost of the
users’ comment speeds calculated
earlier.
5. The stream was considered anomalous
if users were commenting in
greater volume and at greater speed
than would be expected based upon
the stream’s median comment count
and comment speed. For this study,
a value of 3 was used to establish
significance; that is, a mean comment
count at least three times the
median comment count, or a mean
comment speed at most one-third
the median comment speed (in
milliseconds). This was a subjective
definition and can be adjusted for
future studies by editing the config.
ini file.
6. The users of an anomalous stream
were then reviewed. If exhibiting
anomalous behavior of their own,
their comments were recorded for
review by a human reviewer.
a. Anomalous users were defined
as users whose individual
comment count was at least
three times the stream’s median
comment count and who had a
mean comment speed at most
Results
one-third of the stream’s median
comment speed.
The custom Python script analyzed
186,847 users across 128
streams from 14 Twitch channels.
Of these, we found 7,332 anomalous
users across 69 anomalous streams
from 12 of the Twitch channels (see Table
1). An archive of the users and comments
given in Table 1 is available upon
request.
We randomly selected and manually
reviewed 100 anomalous users,
but none were clearly bots. These users
posted comments of varying length and
content, and many responded actively
to other users, suggesting a human was
commenting on the stream.
Discussion
Due to user volume, it was not
feasible to manually review
all identified anomalous users.
The overwhelming majority of the
manually reviewed anomalous users
were engaged in mere chat spam (rapidly
creating or copy-pasting inflammatory,
humorous, or emoji comments).
Surprisingly, neither the Donald Trump
nor Bernie Sanders Twitch channels
had anomalous users. We expected IFO
actors to target voting-age users within
these two channels as the 2020 election
approaches, but the lack of IFOs may
be explained by IFO actors’ ignorance
of the Twitch platform itself. Additionally,
these two channels did not stream
49