Detection of Bias in Media Outlets with Statistical Learning Methods - Text Mining: Classification, Clustering, and Applications - page 40

covered. The definitions of these two distances will be given below. Once we
obtained the distance scores between two news outlets we used the multi-
dimensional scaling (MDS) [see Appendix D ] algorithm to calculate the map
with the optimal positions of the news outlets.
Like in Section 2.4, we ran the matching algorithm on all news outlet pairs.
The time window size was 15 days and the size of the nearest-neighbor list
was 2. The following table shows the number of discovered pairs:

TABLE 2.6:

Number of

discovered pairs.
AJ CNN DN IHT

AJ - 816 447 834

CNN 816 - 1103 2437

DN 447 1103 - 895

IHT 834 2437 895 -

The intersection that we find between the various outlet pairs varies consid-
erably. This can be better seen by calculating the conditional probability of a
story appearing in an outlet given that it appears in another one, reported in
the following table. For example, we can note that more AJ stories feature in
CNN than vice versa ( P ( AJ

AJ )=0 . 3810),
which is perhaps explained by the regional focus of one outlet and the global
focus of the other. Similar relations apply - for example - to the relation
between DN and CNN, or IHT.

| CNN )=0 . 1193 while P ( CNN |

TABLE 2.7: Conditional probabilities of
a story from one news outlet appearing in
another outlet.
P ( X | Y ) AJ CNN DN IHT

AJ 1.0000 0.3810 0.2087 0.3894

CNN 0.1193 1.0000 0.1613 0.3563

DN 0.1526 0.3766 1.0000 0.3056

IHT 0.0865 0.2528 0.0928 1.0000

For the purpose of comparing the vocabulary of news outlets we extracted
the news events which were covered by all four news outlets. We did that by
taking all the news articles that have mates in all other news outlets. The
number of news taken from each outlet can be found in Table 2.8 .
Note that the number of news articles differs slightly between news outlets
with IHT having the largest number of articles and Detroit News the lowest.