In this paper, we propose a method for classifying Twitter accounts
into non-targeting accounts, which post messages to the general
public, and targeting accounts, which post messages to specific
people. For example, an account posting general news information is a
non-targeting account, while an account posting announcements to
members of a specific organization is a targeting account. An account
posting information on very specific minor topic, and an account used
for communication with friends, are also targeting accounts. Our
method finds some properties that are common to most followers of a
given account, and calculate how much a user set with such a
consistency deviates from a random sample from the given universe of
users (e.g., the set of all Twitter users in some country). If it
largely deviates, the account is a targeting account. We use two
types of properties of followers: (1) terms in their metadata and (2)
their followees. The result of our experiment shows that one of our
methods, which computes two scores based on these two types of
properties and combines them using SVM, achieves the accuracy 0.944,
and outperforms the baselines.
microblog;
social network;
target users;
user intention;
target specificity;
target diversity
Published in Proc. of ACM Hypertext, pp.291-296, Halifax, Canada, 2016