Information Technology Reference
In-Depth Information
famous researcher is reachable from a person, then that person can be more produc-
tive. The minimum number of links in the neighbor nodeset of
x
from the cooc net-
work from Japanese web sites
Min
C
(
1
)
G
Jcooc
is also effective, which means
that if a direct neighbor is productive, then
x
will be more productive. The ratio of the
number of edges among neighbors to the number of edges among reachable nodes
from co-project network
Ratio
◦
γ
◦
◦
x
C
(
1
)
C
(
∞
)
s
(
1
)
◦
s
(
1
)
◦
G
pro ject
means
that binding neighbors from all reachable nodes in projects makes the researcher
more productive.
We understand that various features have been shown to be important for
real-world rankings (i.e. target rankings). Some indices correspond to well-known
indices in social network analysis: degree centrality, closeness centrality, and be-
tweenness centrality. Some indices seem new, but their meanings resemble those
of the existing indices. The results support the usefulness of the indices that are
commonly used in the social network literature, and underscore the potential for
additional composition of useful features.
◦
(
Sum
◦
,
Sum
◦
)
◦
x
x
Summary:
Social networks vary according to different relational indices or types
even if they contain the same list of researchers. Researchers have different central-
ity rankings even though they are in relational networks of the same type. Multi-
relational networks have more information than single networks to explain target
rankings. Well-chosen attribute-based features offer good performance for explain-
ing target rankings. However, by combining proposed network-based features, the
prediction results are further improved. Various network-based features have been
shown to be important for real-world rankings (i.e. target rankings), some of which
correspond to well-known indices in social network analysis such as degree cen-
trality, closeness centrality, and betweenness centrality. Some indices seem new, but
their meanings resemble those of existing indices.
6
Related Works
In the context of information retrieval, PageRank [10] and HITS [6] algorithms
can be considered as well known examples for ranking web pages based on the
link structure. Recently, algorithms that are more advanced have been proposed for
learning to rank entities. Although numerous studies of learning-to-rank fields (par-
ticularly targeted on information retrieval) have investigated many attribute-based
ranking functions learned from given preference orders, only a few studies have
concluded that such an impact arises from relations and structures [1, 12]. Fur-
thermore, our model is target-dependent: the important features of relations and
structural embeddedness vary among different tasks.
Relations and structural embeddedness influence the behavior of individuals and
growth and change of the group [13]. Several researchers use network-based fea-
tures for analyses. Backstrom et al. [2] describe analyses of community evolution;
they show some structural features characterizing individuals and positions in the
network. Liben-Nowell et al. [7] elucidate features using network structures in the
link prediction problem. We specifically examine relations and structural features
