Analysis of Content Popularity in Social Bookmarking Systems - Evolving Application Domains of Data Warehousing and Mining - page 249

all terms from the Digg stories, it was possible to
create the two-way contingency matrix of Table 4
for each term t . According to this, A is the number
of popular stories containing term t , B the number
of non-popular stories containing t , and C and D
are the number of popular and non-popular stories
respectively that don't contain term t .
Then, the χ 2 statistic is calculated based on the
following equation:

words and numeric strings were also filtered out
of the feature selection process. In order to keep
the experiments simple, no stemming or other
term normalization was carried out. Table 5 lists
the top 30 terms of this story set along with their
χ 2 scores. Although not very informative to the
human inspector, these keywords can be consid-
ered as the most appropriate (from a text feature
perspective) for use in making the distinction
between Popular and Non-popular stories.
In order to get a more fine-grained view of
such keywords per topic, we also calculate the χ 2
scores on independent corpora that contain stories
only from particular topics. Table 6 provides such
a topic-specific χ 2 -based ranking of terms.
After ranking the terms of each corpus based
on their class separation ability, it is possible
to select the top K of them and use them in an
automatic text classification scheme. Table 7
presents the results achieved by such a classifi-
cation scheme, i.e. the success of predicting the
popularity of Digg stories, where three classifiers
are compared, namely a Naïve Bayes classifier, an
SVM and a C4.5 decision tree. The dataset used
consists of 50,000 randomly selected stories and
the performance metrics were calculated by use of
10-fold cross validation (i.e. repeated splitting of
the dataset to 10 parts and usage of nine of them
for training and one of them for testing).

= × × -×
+× +× +× +

NADCB
AC BD ABCD

( )

2

c 2 () t

( ) ( ) ( ) ( )

(8)

The χ 2 statistic naturally takes a value of zero if
term t and the class of Popular stories are indepen-
dent. The measure is only problematic when any
of the contingency table cells is lightly populated,
which is the case for low-frequency terms. For
that reason, we filter low-frequency terms prior
to the calculation of Equation 8. In addition, stop

Table 4. Two-way contingency table of term t

Popular Non-popular

Term exists A B

Term doesn't exist

C D

Table 5. Top 30 text features based on χ 2 statistic

# Term χ 2 (10 6 ) # Term χ 2 (10 6 ) # Term χ 2 (10 6 )

1 see 61.5 11 seen 7.7 21 news 4.2

2 drive 60.4 12 nintendo 7.6 22 making 4.1

3 japanese 28.5 13 program 5.6 23 breaking 4.0

4 video 17.2 14 way 5.4 24 amazing 3.5

5 google 12.9 15 gets 5.2 25 say 3.4

6 long 11.7 16 computer 4.9 26 coolest 3.4

7 cool 11.0 17 need 4.8 27 release 3.3

8 term 9.9 18 want 4.7 28 right .9

9 look 8.8 9 play 4.7 29 xbox 2.8

10 high 7.9 20 job 4.3 30 looks .5