We then probe our model’s generality, by transfering
its learned representations to the task of personalized
document recommendation: for each of
M users, given N previous positive interactions
with documents (likes, clicks, etc.), predict the
N + 1’th document the user will positively interact
with. To perform well on this task, the representation
should capture the user’s interest in
textual content. We find representations trained
on hashtag prediction outperform representations
from unsupervised learning, and that our convolution