Information Technology Reference
In-Depth Information
and
r
k
denotes
k
-th element of the vector
r
∗
and means the target ranking score of
entity
v
k
), the goal is to learn a ranking model based on these networks.
First, as a baseline approach, we follow the intuitive idea of simply using the
approach from SNA (i.e. centrality) to learn ranking. Then we propose a more sys-
tematic algorithm that generates various network features for individuals from social
networks.
4.1
Baseline Model
Based on the intuitive approach, we first overview commonly used indices in social
network analysis and complex network studies. Given a set of social networks, we
rank entities on these networks using different network centrality indices. We des-
ignate these rankings as
network rankings
because they are calculated directly from
relational networks.
To address the question of what kind of relation is most important for entities,
we intuitively compare rankings resulting from relations of various types. Although
simple, it can be considered as an implicit step of social network analysis given a
set of relational networks. We merely choose the type of relation that maximally
explains the given ranking. We rank the relational network of each type; then we
compare the
network ranking
with the
target ranking
. Intuitively, if the correlation
to the network ranking
r
i
is high, then the relation
i
represents important influences
among entities for the given target ranking. Therefore, this model is designed to
determine an optimal relation
i
from a set of relations:
i
r
∗
)
,
=
argmax
i
∈{
1
,...,
m
}
Cor
(
r
i
,
(1)
For different relational networks with different centrality indices, the network rank-
ing from
i
-th network with
j
-th centrality ranking can be represented as
r
i
,
j
(
R
n
),
∈
where
i
. Therefore, the first method can be extended
simply to find a pair of optimal parameters
∈{
1
,...,
m
}
,and
j
∈{
1
,...,
s
}
i
j
(i.e. the
i
-th network by
j
-th
centrality rankings) that maximizes the coefficient between network rankings with
a target ranking.
<
,
>
i
j
r
∗
)
,
<
,
>
=
argmax
Cor
(
r
i
,
j
,
(2)
i
∈{
1
,...,
m
}
j
∈{
1
,...,
s
}
4.2
Network Combination Model
Many centrality approaches related to ranking network entities specifically exam-
ine graphs with a single link type. However, multiple social networks exist in the
real world, each representing a particular relation type; each of which might be in-
tegrated to play a distinct role in a particular task. We combine several extracted
multiple social networks into one network and designate such a social network
as a
combined-relational network
(denoted as
G
c
(
V
,
E
c
)
). Our target is using a
