Data Representation
We represent each article as a vector of features corresponding to the words in
the article's abstract.
We apply the Porter stemming algorithm (Porter, 1980)
to the abstract text,
count the use of each word in each abstract using
MALLET (McCallum, 2002),
and limit the vectors to words that appear at least three times in the corpus.
This reduces the vocabulary size in our training data set
by 47% (17,060 words appear once, 6,534 words appear twice,
and we retain a vocabulary of 26,364 words that appear at least three times).
Finally, we represent each word by the value log(1+ni),
where ni is the number of
times word i appears in an abstract, and we normalize the vector
for each article to have unit Euclidean length,
i.e. to lie on the unit sphere.
We do not use information associated with the articles (journal names, author names, etc.)
other than the abstract text. We verify that the vocabulary
does not contain any obvious "leaker" words,
i.e. terms that discriminate between the training subsets
but are not valid predictors for future articles.
In particular, the trained models do not use article dates,
directly or indirectly, to make predictions.
