<s>
In	O
network	O
theory	O
,	O
link	B-Application
prediction	I-Application
is	O
the	O
problem	O
of	O
predicting	O
the	O
existence	O
of	O
a	O
link	O
between	O
two	O
entities	O
in	O
a	O
network	O
.	O
</s>
<s>
Examples	O
of	O
link	B-Application
prediction	I-Application
include	O
predicting	O
friendship	O
links	O
among	O
users	O
in	O
a	O
social	O
network	O
,	O
predicting	O
co-authorship	O
links	O
in	O
a	O
citation	O
network	O
,	O
and	O
predicting	O
interactions	O
between	O
genes	O
and	O
proteins	O
in	O
a	O
biological	O
network	O
.	O
</s>
<s>
Link	B-Application
prediction	I-Application
can	O
also	O
have	O
a	O
temporal	O
aspect	O
,	O
where	O
,	O
given	O
a	O
snapshot	O
of	O
the	O
set	O
of	O
links	O
at	O
time	O
,	O
the	O
goal	O
is	O
to	O
predict	O
the	O
links	O
at	O
time	O
.	O
</s>
<s>
Link	B-Application
prediction	I-Application
is	O
widely	O
applicable	O
.	O
</s>
<s>
In	O
e-commerce	O
,	O
link	B-Application
prediction	I-Application
is	O
often	O
a	O
subtask	O
for	O
recommending	O
items	O
to	O
users	O
.	O
</s>
<s>
The	O
goal	O
of	O
link	B-Application
prediction	I-Application
is	O
to	O
identify	O
the	O
unobserved	O
true	O
links	O
.	O
</s>
<s>
In	O
the	O
binary	O
classification	O
formulation	O
of	O
the	O
link	B-Application
prediction	I-Application
task	O
the	O
potential	O
links	O
are	O
classified	O
as	O
either	O
true	O
links	O
or	O
false	O
links	O
.	O
</s>
<s>
Link	B-Application
prediction	I-Application
approaches	O
for	O
this	O
setting	O
learn	O
a	O
classifier	O
that	O
maps	O
links	O
in	O
to	O
positive	O
and	O
negative	O
labels	O
i.e.	O
</s>
<s>
Link	B-Application
prediction	I-Application
approaches	O
for	O
this	O
setting	O
learn	O
a	O
model	O
that	O
maps	O
links	O
in	O
to	O
a	O
probability	O
i.e.	O
</s>
<s>
Structured	O
prediction	O
approaches	O
capture	O
the	O
correlation	O
between	O
potential	O
links	O
by	O
formulating	O
the	O
task	O
as	O
a	O
collective	O
link	B-Application
prediction	I-Application
task	O
.	O
</s>
<s>
Collective	O
link	B-Application
prediction	I-Application
approaches	O
learn	O
a	O
model	O
that	O
jointly	O
identify	O
all	O
the	O
true	O
links	O
among	O
the	O
set	O
of	O
potential	O
links	O
.	O
</s>
<s>
Link	B-Application
prediction	I-Application
task	O
can	O
also	O
be	O
formulated	O
as	O
an	O
instance	O
of	O
missing	O
value	O
estimation	O
task	O
.	O
</s>
<s>
Here	O
,	O
the	O
graph	O
is	O
represented	O
as	O
an	O
adjacency	B-Algorithm
matrix	I-Algorithm
with	O
missing	O
values	O
.	O
</s>
<s>
Matrix	B-Application
factorization	I-Application
based	O
methods	O
commonly	O
use	O
this	O
formulation	O
.	O
</s>
<s>
The	O
task	O
of	O
link	B-Application
prediction	I-Application
has	O
attracted	O
attention	O
from	O
several	O
research	O
communities	O
ranging	O
from	O
statistics	O
and	O
network	O
science	O
to	O
machine	O
learning	O
and	O
data	B-Application
mining	I-Application
.	O
</s>
<s>
For	O
social	O
networks	O
,	O
Liben-Nowell	O
and	O
Kleinberg	O
proposed	O
a	O
link	B-Application
prediction	I-Application
models	O
based	O
on	O
different	O
graph	O
proximity	O
measures	O
.	O
</s>
<s>
Several	O
statistical	O
models	O
have	O
been	O
proposed	O
for	O
link	B-Application
prediction	I-Application
by	O
the	O
machine	O
learning	O
and	O
data	B-Application
mining	I-Application
community	O
.	O
</s>
<s>
Several	O
models	O
based	O
on	O
directed	O
graphical	B-Algorithm
models	I-Algorithm
for	O
collective	O
link	B-Application
prediction	I-Application
have	O
been	O
proposed	O
by	O
Getoor	O
.	O
</s>
<s>
With	O
the	O
advent	O
of	O
deep	B-Algorithm
learning	I-Algorithm
,	O
several	O
graph	B-Algorithm
embedding	I-Algorithm
based	O
approaches	O
for	O
link	B-Application
prediction	I-Application
have	O
also	O
been	O
proposed	O
.	O
</s>
<s>
For	O
more	O
information	O
on	O
link	B-Application
prediction	I-Application
refer	O
to	O
the	O
survey	O
by	O
Getoor	O
et	O
al	O
.	O
</s>
<s>
Several	O
link	O
predication	O
approaches	O
have	O
been	O
proposed	O
including	O
unsupervised	O
approaches	O
such	O
as	O
similarity	O
measures	O
computed	O
on	O
the	O
entity	O
attributes	O
,	O
random	O
walk	O
and	O
matrix	B-Application
factorization	I-Application
based	O
approaches	O
,	O
and	O
supervised	O
approaches	O
based	O
on	O
graphical	B-Algorithm
models	I-Algorithm
and	O
deep	B-Algorithm
learning	I-Algorithm
.	O
</s>
<s>
Link	B-Application
prediction	I-Application
approaches	O
can	O
be	O
divided	O
into	O
two	O
broad	O
categories	O
based	O
on	O
the	O
type	O
of	O
the	O
underlying	O
network	O
:	O
(	O
1	O
)	O
link	B-Application
prediction	I-Application
approaches	O
for	O
homogeneous	O
networks	O
(	O
2	O
)	O
link	B-Application
prediction	I-Application
approaches	O
for	O
heterogeneous	O
networks	O
.	O
</s>
<s>
This	O
is	O
a	O
common	O
approach	O
to	O
link	B-Application
prediction	I-Application
that	O
computes	O
the	O
number	O
of	O
common	O
neighbors	O
.	O
</s>
<s>
Let	O
A	O
be	O
the	O
adjacency	B-Algorithm
matrix	I-Algorithm
of	O
a	O
network	O
under	O
consideration	O
.	O
</s>
<s>
Graph	B-Algorithm
embeddings	I-Algorithm
also	O
offer	O
a	O
convenient	O
way	O
to	O
predict	O
links	O
.	O
</s>
<s>
Graph	B-Algorithm
embedding	I-Algorithm
algorithms	O
,	O
such	O
as	O
Node2vec	B-General_Concept
,	O
learn	O
an	O
embedding	O
space	O
in	O
which	O
neighboring	O
nodes	O
are	O
represented	O
by	O
vectors	O
so	O
that	O
vector	O
similarity	O
measures	O
,	O
such	O
as	O
dot	O
product	O
similarity	O
,	O
or	O
euclidean	O
distance	O
,	O
hold	O
in	O
the	O
embedding	O
space	O
.	O
</s>
<s>
MLNs	O
are	O
able	O
to	O
incorporate	O
both	O
local	O
and	O
relational	O
rules	O
for	O
the	O
purpose	O
of	O
link	B-Application
prediction	I-Application
.	O
</s>
<s>
R-Models	O
(	O
RMLs	O
)	O
is	O
a	O
neural	O
network	O
model	O
created	O
to	O
provide	O
a	O
deep	B-Algorithm
learning	I-Algorithm
approach	O
to	O
the	O
link	O
weight	O
prediction	O
problem	O
.	O
</s>
<s>
Link	B-Application
prediction	I-Application
has	O
found	O
varied	O
uses	O
,	O
but	O
any	O
domain	O
in	O
which	O
entities	O
interact	O
in	O
a	O
structures	O
way	O
can	O
benefit	O
from	O
link	B-Application
prediction	I-Application
.	O
</s>
<s>
A	O
common	O
applications	O
of	O
link	B-Application
prediction	I-Application
is	O
improving	O
similarity	O
measures	O
for	O
collaborative	B-Algorithm
filtering	I-Algorithm
approaches	O
to	O
recommendation	O
.	O
</s>
<s>
Link	B-Application
prediction	I-Application
is	O
also	O
frequently	O
used	O
in	O
social	O
networks	O
to	O
suggest	O
friends	O
to	O
users	O
.	O
</s>
<s>
In	O
biology	O
,	O
link	B-Application
prediction	I-Application
has	O
been	O
used	O
to	O
predict	O
interactions	O
between	O
proteins	O
in	O
protein-protein	O
interaction	O
networks	O
.	O
</s>
<s>
Link	B-Application
prediction	I-Application
has	O
also	O
been	O
used	O
to	O
infer	O
interactions	O
between	O
drugs	O
and	O
targets	O
using	O
link	B-Application
prediction	I-Application
Another	O
application	O
is	O
found	O
in	O
collaboration	O
prediction	O
in	O
scientific	O
co-authorship	O
networks	O
.	O
</s>
<s>
Entity	B-General_Concept
resolution	I-General_Concept
,	O
also	O
known	O
as	O
deduplication	O
,	O
commonly	O
uses	O
link	B-Application
prediction	I-Application
to	O
predict	O
whether	O
two	O
entities	O
in	O
a	O
network	O
are	O
references	O
to	O
the	O
same	O
physical	O
entity	O
.	O
</s>
<s>
Some	O
authors	O
have	O
used	O
context	O
information	O
in	O
network	O
structured	O
domains	O
to	O
improve	O
entity	B-General_Concept
resolution	I-General_Concept
.	O
</s>
<s>
Link	B-Application
prediction	I-Application
in	O
the	O
context	O
of	O
network	O
effects	O
has	O
been	O
used	O
to	O
analyze	O
tendencies	O
to	O
spread	O
across	O
networks	O
and	O
can	O
be	O
used	O
to	O
improve	O
marketing	O
strategies	O
,	O
in	O
particular	O
viral	O
marketing	O
.	O
</s>
