,
10 tweets,
5 min read
Read on Twitter
We did a couple of experiments this week with techniques we haven't used frequently (sentiment analysis and clustering). Here are some early thoughts. . .
cc: @ZellaQuixote
cc: @ZellaQuixote
We used VADER for the sentiment analysis portion. This tool scores individual words on a negative/positive emotional scale, and then combines those scores to give an overall score for a piece of text ranging from -1 (most negative) to 1 (most positive).
comp.social.gatech.edu/papers/icwsm14…
comp.social.gatech.edu/papers/icwsm14…
These charts show the results of running a few accounts' tweets (retweets excluded) through VADER sentiment analysis. Apparently we're both a bit more positive than negative. @PamelaGeller, on the other hand. . 
Next, we tried sentiment analysis on the #Novichok tweets from the month following the Skripal poisoning. The tweets skew decidedly negative, but without context it's impossible to tell what they're being negative about.
Retweet network for #Novichok, colored by average sentiment score of each account's Novichok tweets. The structure of the graph (like-minded accounts group together) proves ultimately more useful than the sentiment scores for determining the stance of a given user.
(old #Novichok thread here)
Next up: k-means clustering. This is a technique for splitting a set of data points into smaller groups such that the data points in each group are similar to one another. We'll attempt to use it to pick out groups of bots from a larger set.
en.wikipedia.org/wiki/K-means_c…
en.wikipedia.org/wiki/K-means_c…
In this thread we found that 5576 of the accounts linking to powerlineblog(dot)com were likely automated. We first tried clustering based on apps used to tweet and news services linked, with mixed success.
Clustering the accounts based on tweet schedule was more effective, as shown here. There are still some errors, but it's sufficient for the human eye to look at the results and start picking out groups of accounts that are doing the same thing.
Important takeaway from both experiments: while there is value in both approaches, machine-learning/AI techniques only go so far as social media analysis tools without human-supplied observation and context. (This is among the reasons we no longer run an AI account-scoring tool.)
