<s>
Cohen	B-General_Concept
's	I-General_Concept
kappa	I-General_Concept
coefficient	O
( κ	O
,	O
lowercase	O
Greek	O
kappa	O
)	O
is	O
a	O
statistic	O
that	O
is	O
used	O
to	O
measure	O
inter-rater	O
reliability	O
(	O
and	O
also	O
intra-rater	O
reliability	O
)	O
for	O
qualitative	O
(	O
categorical	O
)	O
items	O
.	O
</s>
<s>
There	O
is	O
controversy	O
surrounding	O
Cohen	B-General_Concept
's	I-General_Concept
kappa	I-General_Concept
due	O
to	O
the	O
difficulty	O
in	O
interpreting	O
indices	O
of	O
agreement	O
.	O
</s>
<s>
Cohen	B-General_Concept
's	I-General_Concept
kappa	I-General_Concept
measures	O
the	O
agreement	O
between	O
two	O
raters	O
who	O
each	O
classify	O
N	O
items	O
into	O
C	O
mutually	O
exclusive	O
categories	O
.	O
</s>
<s>
In	O
the	O
traditional	O
2	O
×	O
2	O
confusion	B-General_Concept
matrix	I-General_Concept
employed	O
in	O
machine	O
learning	O
and	O
statistics	O
to	O
evaluate	O
binary	B-General_Concept
classifications	I-General_Concept
,	O
the	O
Cohen	B-General_Concept
's	I-General_Concept
Kappa	I-General_Concept
formula	O
can	O
be	O
written	O
as	O
:	O
</s>
<s>
In	O
this	O
case	O
,	O
Cohen	B-General_Concept
's	I-General_Concept
Kappa	I-General_Concept
is	O
equivalent	O
to	O
the	O
Heidke	O
skill	O
score	O
known	O
in	O
Meteorology	O
.	O
</s>
<s>
So	O
now	O
applying	O
our	O
formula	O
for	O
Cohen	B-General_Concept
's	I-General_Concept
Kappa	I-General_Concept
we	O
get	O
:	O
</s>
<s>
A	O
case	O
sometimes	O
considered	O
to	O
be	O
a	O
problem	O
with	O
Cohen	B-General_Concept
's	I-General_Concept
Kappa	I-General_Concept
occurs	O
when	O
comparing	O
the	O
Kappa	O
calculated	O
for	O
two	O
pairs	O
of	O
raters	O
with	O
the	O
two	O
raters	O
in	O
each	O
pair	O
having	O
the	O
same	O
percentage	O
agreement	O
but	O
one	O
pair	O
give	O
a	O
similar	O
number	O
of	O
ratings	O
in	O
each	O
class	O
while	O
the	O
other	O
pair	O
give	O
a	O
very	O
different	O
number	O
of	O
ratings	O
in	O
each	O
class	O
.	O
</s>
<s>
For	O
instance	O
,	O
in	O
the	O
following	O
two	O
cases	O
there	O
is	O
equal	O
agreement	O
between	O
A	O
and	O
B	O
(	O
60	O
out	O
of	O
100	O
in	O
both	O
cases	O
)	O
in	O
terms	O
of	O
agreement	O
in	O
each	O
class	O
,	O
so	O
we	O
would	O
expect	O
the	O
relative	O
values	O
of	O
Cohen	B-General_Concept
's	I-General_Concept
Kappa	I-General_Concept
to	O
reflect	O
this	O
.	O
</s>
<s>
However	O
,	O
calculating	O
Cohen	B-General_Concept
's	I-General_Concept
Kappa	I-General_Concept
for	O
each	O
:	O
</s>
<s>
P-value	B-General_Concept
for	O
kappa	O
is	O
rarely	O
reported	O
,	O
probably	O
because	O
even	O
relatively	O
low	O
values	O
of	O
kappa	O
can	O
nonetheless	O
be	O
significantly	O
different	O
from	O
zero	O
but	O
not	O
of	O
sufficient	O
magnitude	O
to	O
satisfy	O
investigators	O
.	O
</s>
<s>
(	O
and	O
the	O
CI	O
in	O
general	O
)	O
may	O
also	O
be	O
estimated	O
using	O
bootstrap	B-Application
methods	I-Application
.	O
</s>
<s>
The	O
so-called	O
chance	O
adjustment	O
of	O
kappa	B-General_Concept
statistics	I-General_Concept
supposes	O
that	O
,	O
when	O
not	O
completely	O
certain	O
,	O
raters	O
simply	O
guess	O
—	O
a	O
very	O
unrealistic	O
scenario	O
.	O
</s>
<s>
Moreover	O
,	O
some	O
works	O
have	O
shown	O
how	O
kappa	B-General_Concept
statistics	I-General_Concept
can	O
lead	O
to	O
a	O
wrong	O
conclusion	O
for	O
unbalanced	O
data	O
.	O
</s>
<s>
Cohen	B-General_Concept
's	I-General_Concept
kappa	I-General_Concept
and	O
Scott	O
's	O
pi	O
differ	O
in	O
terms	O
of	O
how	O
is	O
calculated	O
.	O
</s>
<s>
Note	O
that	O
Cohen	B-General_Concept
's	I-General_Concept
kappa	I-General_Concept
measures	O
agreement	O
between	O
two	O
raters	O
only	O
.	O
</s>
<s>
The	O
Fleiss	O
kappa	O
,	O
however	O
,	O
is	O
a	O
multi-rater	O
generalization	O
of	O
Scott	O
's	O
pi	O
statistic	O
,	O
not	O
Cohen	B-General_Concept
's	I-General_Concept
kappa	I-General_Concept
.	O
</s>
<s>
Kappa	O
is	O
also	O
used	O
to	O
compare	O
performance	O
in	O
machine	O
learning	O
,	O
but	O
the	O
directional	O
version	O
known	O
as	O
Informedness	B-General_Concept
or	O
Youden	B-General_Concept
's	I-General_Concept
J	I-General_Concept
statistic	I-General_Concept
is	O
argued	O
to	O
be	O
more	O
appropriate	O
for	O
supervised	O
learning	O
.	O
</s>
