<s>
Expectation	B-General_Concept
propagation	I-General_Concept
(	O
EP	O
)	O
is	O
a	O
technique	O
in	O
Bayesian	O
machine	O
learning	O
.	O
</s>
<s>
It	O
uses	O
an	O
iterative	B-Algorithm
approach	O
that	O
uses	O
the	O
factorization	O
structure	O
of	O
the	O
target	O
distribution	O
.	O
</s>
<s>
Expectation	B-General_Concept
propagation	I-General_Concept
achieves	O
this	O
approximation	O
by	O
minimizing	O
the	O
Kullback-Leibler	O
divergence	O
.	O
</s>
<s>
Expectation	B-General_Concept
propagation	I-General_Concept
via	O
moment	O
matching	O
plays	O
a	O
vital	O
role	O
in	O
approximation	O
for	O
indicator	O
functions	O
that	O
appear	O
when	O
deriving	O
the	O
message	O
passing	O
equations	O
for	O
TrueSkill	B-Device
.	O
</s>
