Collective Classification for Text Classification - Text Mining: Classification, Clustering, and Applications - page 59

3.5 Learning the Classifiers
One aspect of the collective classification problem that we have not dis-
cussed so far is how to learn the various classifiers described in the previous
sections. Learning refers to the problem of determining the parameter val-
ues for the local classifier, in the case of ICA and GS, and the values in the
clique potentials, in the case of LBP and MF, which can then be subsequently
used to classify unseen test data. For all our experiments, we learned the pa-
rameter values from fully labeled datasets using gradient-based optimization
approaches. Unfortunately, a full treatment of this subject is not possible
within this article and we refer the interested reader to various other works
that discuss this in more depth such as (34), (31), (32).
3.6 Experimental Comparison
In our evaluation, we compared the four collective classification algorithms
(CC) discussed in the previous sections and a content-only classifier (CO),
which does not take the link structure into account, along with two choices
of local classifiers on document classification tasks. The two local classifiers
we tried were na ıve Bayes (NB) and Logistic Regression (LR). This gave us
8 different classifiers: CO with NB, CO with LR, ICA with NB, ICA with
LR, GS with NB, GS with LR, MF and LBP. The datasets we used for the
experiments included both real-world and synthetic datasets.
3.6.1 Features Used
For CO classifiers, we used the words in the documents for observed at-
tributes. In particular, we used a binary value to indicate whether or not
a word appears in the document. In ICA and GS, we used the same local
attributes (i.e., words) followed by count aggregation to count the number of
each label value in a node's neighborhood. Finally, for LBP and MF, we used
pairwise Markov Random Fields with clique potentials defined on the edges
and unobserved nodes in the network.
3.6.2 Real-World Datasets
We experimented with three real-world datasets: Cora and CiteSeer (two
bibliographic datasets), and WebKB (a hypertext dataset). For the WebKB
experiments, we only considered documents which link to or are linked to by
at least one other webpage in the corpus. This gave us a corpus of size 877