4. Contribu/ons
Propose
OMNI-‐Prop:
a
node
classifica/on
algorithm
• Seamless
and
Accurate
– good
accuracy
on
arbitrary
label
correla/on
• Fast
– each
itera/on
is
linear
on
input
graph
size
– convergence
guarantee
• (Quasi-‐parameter
free)
-‐
omiZed
in
this
talk
for
brevity
– Just
one
parameter
with
default
value
1
– No
parameter
to
tune
15/01/29
Yuto
Yamaguchi
-‐
AAAI2015
4
6. Basic
Idea
15/01/29
Yuto
Yamaguchi
-‐
AAAI2015
6
If most of the neighbors of a node have the same label,
then the rest also have the same label.
?
Most
neighbors
are
the
same
à
the
rest
is
also
the
same
Neighbors
have
different
labels
à
say
nothing
?
?
7. How
it
works?
15/01/29
Yuto
Yamaguchi
-‐
AAAI2015
7
• sij:
How
likely
node
i
has
label
j
• tij:
How
likely
the
neighbors
of
node
i
have
label
j
Calculate
two
variables
recursively
male
male
male
unknown
male
male
male
male?
most
friends
are
males!
I
am
a
male
s-‐propaga5on
t-‐propaga5on
s
s
s
s
ß
aggrega/on
of
t
ß
aggrega/on
of
s
t
t
t
t
you
are
probably
males
see
paper
for
details
?
?
9. Complexity
and
Convergence
15/01/29
Yuto
Yamaguchi
-‐
AAAI2015
9
*
K:
#
labels
N:
#
nodes
M:
#
edges
[Theorem 1 - complexity]
The time complexity of each iteration
of OMNI-Prop is O(K(N+M))
[Theorem 2 - convergence]
OMNI-Prop always converges on arbitrary graphs
10. Theore/cal
connec/on
to
SSL
15/01/29
Yuto
Yamaguchi
-‐
AAAI2015
10
Label
Propaga/on
[Zhu+,
2003]
Original
graph
Twin
graph
[Theorem 3 - equivalence]
The special case of OMNI-Prop is equivalent
to LP on twin graph
14. Summary
• Proposed
OMNI-‐Prop
– Seamless
NL
on
arbitrary
label
correla/on
– Fast
– (Quasi-‐parameter
free)
• Theore/cally
– Linear
on
input
size
for
each
itera/on
– Always
converges
on
arbitrary
graphs
– special
case
=
LP
• Experimentally
– Almost
always
wins
on
all
5
datasets
15/01/29
Yuto
Yamaguchi
-‐
AAAI2015
14