Fig. 5 and Table 6 demonstrate that our approach has the
ability to detect potential student problems from tweets.
Yet, we are not trying to make the claim that these users
shown are definitely at-risk students, since most of these
users only posted less than 10 percent of problems among
all tweets they have posted. The trained detector can be
applied as a monitoring mechanism in the long run to identify
severe cases of at-risk students. For example, a future
student may post a large number of tweets and more than
90 percent of them are about study problems or negative
emotions. This case can be detected using the detector presented
here. Our Purdue data set only lasts a little over two
months. However, severe cases may only appear once in a
while depending on the institutional atmosphere. Further
decisions need to be made about what counts as severe
cases, to what extent does intervention is needed, how to
protect students’ privacy, and how comfortable they are
about these interventions.