Information Technology Reference
In-Depth Information
and predict the rankings. We can obtain a 0
448 correlation coefficient between pre-
dicted rankings and target rankings, which seems to be readily explainable: famous
researchers are also famous on the web sites. Subsequently, we integrate the pro-
posed network-based features obtained from each type of single network as well as
multi-relational networks among researchers to train and predict the rankings. The
co-occurrence-based networks
G
Ecooc
,
G
Eoverla p
,
G
Joverla p
(especially on English-
language web sites) appear to be a better explanation of target ranking of
Paper
than
the co-affiliation network
G
af filiation
or the co-project network
G
pro jext
. Using fea-
tures from multi-relational networks
G
ALL
, the prediction results are better than for
any other single-relational network. Furthermore, when we combine network-based
features with attribute-based features to learn the model, the results outperform each
using attribute-based features only and network-based features only.
.
5.3
Detailed Analysis of Useful Features
We use network-based features separately for training. We expect the target rank-
ings to clarify their usefulness. Leaving out one feature, the others are used to train
and predict the rankings to evaluate network-based features. In fact, the
k
-th feature
is a useful feature for explaining the target ranking if the result worsens much when
leaving out the
k
feature. Table 5 presents the effective features for the target rank-
ing of
Paper
in the researcher field. For example, the maximum number of links
in the reachable nodeset of
x
from the cooc network from English-language web
sites
Max
C
(
∞
)
◦
γ
◦
◦
G
Ecooc
is effective for the target ranking, which means that if a
x
Ta b l e 5
Effective features in various networks for
Paper
among researchers
Top Effective Features for
Paper
Max
◦
γ
◦
C
(
∞
)
1
◦
G
Ecooc
x
Min
◦
γ
◦
C
(
1
)
◦
G
Jcooc
2
x
Avg
◦
γ
◦
C
(
∞
)
3
◦
G
Eoverlap
x
Max
◦
t
◦
C
(
∞
)
4
◦
G
Joverlap
x
Avg
◦
u
x
◦
C
(
1
)
5
◦
G
Eoverlap
x
Min
◦
γ
◦
C
(
1
)
6
◦
G
Eoverlap
x
Min
◦
γ
◦
C
(
∞
)
x
7
◦
G
Jcooc
Ratio
◦
(
Sum
◦
s
(
1
)
◦
C
(
1
)
,
Sum
◦
s
(
1
)
◦
C
(
∞
)
8
)
◦
G
pro ject
x
x
Avg
◦
γ
◦
C
(
1
)
9
◦
G
Joverlap
x
C
(
1
)
x
10
Min
G
Ecooc
11
Ratio
◦
(
Sum
◦
s
(
1
)
◦
C
(
1
)
◦
γ
◦
◦
,
Sum
◦
s
(
1
)
◦
C
(
∞
)
)
◦
G
Ecooc
x
x
12
Ratio
◦
(
Sum
◦
u
x
◦
C
(
1
)
,
Sum
◦
u
x
◦
C
(
∞
)
)
◦
G
Ecooc
x
x
C
(
1
)
x
13
Min
G
Jcooc
14
Ratio
◦
(
Avg
◦
u
x
◦
C
(
1
)
◦
u
x
◦
◦
,
Avg
◦
u
x
◦
C
(
∞
)
)
◦
G
Jcooc
x
x
15
Min
◦
γ
◦
C
(
∞
)
◦
G
Joverlap
x
