Comparison to Production System We also
compare to a proprietary system in production in
Facebook for hashtag prediction. It trains a logistic
regression model for every hashtag, using
a bag of unigrams, bigrams, and trigrams as the input features. Unlike the other models we consider
here, this baseline has been trained using a
set of approximately 10 million posts. Engineering
constraints prevent measuring mean rank performance.
We present it here as a serious effort
at solving the same problem from outside the embedding
paradigm. On the people dataset this system
achieves 3.47% P@1 and 5.33% R@10. On
the pages dataset it obtains 5.97% P@1 and 6.30%
R@10. It is thus outperformed by our method.
However, we note the differences in experimental
setting mean this comparison is perhaps not
completely fair (different training sets). We expect
performance of linear models such as this to be
similar to WSABIE as that has been in the case in
other datasets (Gupta et al., 2014), but at the cost
of more memory usage. Note that models like logistic
regression and SVM do not scale well if you
have millions of hashtags, which we could handle
in our models.
4.3 Personalized document recommendation
To investigate the generality of these learned representations,
we apply them to the task of recommending
documents to users based on the user’s
interaction history. The data for this task comprise
anonymized day-long interaction histories for a
tiny subset of people on a popular social network-
Method dim P@1 R@10 R@50
Word2Vec 256 0.75% 1.96% 3.82%
BoW - 1.36% 4.29% 8.03%
WSABIE 64 0.98% 3.14% 6.65%
WSABIE 128 1.02% 3.30% 6.71%
WSABIE 256 1.01% 2.98% 5.99%
WSABIE 512 1.01% 2.76% 5.19%
#TAGSPACE 64 1.27% 4.56% 9.64%
#TAGSPACE 128 1.48% 4.74% 9.96%
#TAGSPACE 256 1.66% 5.29% 10.69%
WSABIE+ BoW 64 1.61% 4.83% 9.00%
#TAGSPACE+ BoW 64 1.80% 5.90% 11.22%
#TAGSPACE+ BoW 256 1.92% 6.15% 11.53%
Table 5: Document recommendation task results.
ing service. For each of the 34 thousand people
considered, we collected the text of between 5 and
167 posts that she has expressed previous positive
interactions with (likes, clicks, etc.). Given the
person’s trailing n