<s>
In	O
machine	O
learning	O
and	O
statistics	O
,	O
feature	B-General_Concept
selection	I-General_Concept
,	O
also	O
known	O
as	O
variable	B-General_Concept
selection	I-General_Concept
,	O
attribute	O
selection	O
or	O
variable	O
subset	O
selection	O
,	O
is	O
the	O
process	O
of	O
selecting	O
a	O
subset	O
of	O
relevant	O
features	B-Algorithm
(	O
variables	O
,	O
predictors	O
)	O
for	O
use	O
in	O
model	O
construction	O
.	O
</s>
<s>
Feature	B-General_Concept
selection	I-General_Concept
techniques	O
are	O
used	O
for	O
several	O
reasons	O
:	O
</s>
<s>
to	O
avoid	O
the	O
curse	B-Algorithm
of	I-Algorithm
dimensionality	I-Algorithm
,	O
</s>
<s>
The	O
central	O
premise	O
when	O
using	O
a	O
feature	B-General_Concept
selection	I-General_Concept
technique	O
is	O
that	O
the	O
data	O
contains	O
some	O
features	B-Algorithm
that	O
are	O
either	O
redundant	O
or	O
irrelevant	O
,	O
and	O
can	O
thus	O
be	O
removed	O
without	O
incurring	O
much	O
loss	O
of	O
information	O
.	O
</s>
<s>
Feature	B-General_Concept
selection	I-General_Concept
techniques	O
should	O
be	O
distinguished	O
from	O
feature	B-Algorithm
extraction	I-Algorithm
.	O
</s>
<s>
Feature	B-Algorithm
extraction	I-Algorithm
creates	O
new	O
features	B-Algorithm
from	O
functions	O
of	O
the	O
original	O
features	B-Algorithm
,	O
whereas	O
feature	B-General_Concept
selection	I-General_Concept
returns	O
a	O
subset	O
of	O
the	O
features	B-Algorithm
.	O
</s>
<s>
Feature	B-General_Concept
selection	I-General_Concept
techniques	O
are	O
often	O
used	O
in	O
domains	O
where	O
there	O
are	O
many	O
features	B-Algorithm
and	O
comparatively	O
few	O
samples	O
(	O
or	O
data	O
points	O
)	O
.	O
</s>
<s>
Archetypal	O
cases	O
for	O
the	O
application	O
of	O
feature	B-General_Concept
selection	I-General_Concept
include	O
the	O
analysis	O
of	O
written	O
texts	O
and	O
DNA	O
microarray	O
data	O
,	O
where	O
there	O
are	O
many	O
thousands	O
of	O
features	B-Algorithm
,	O
and	O
a	O
few	O
tens	O
to	O
hundreds	O
of	O
samples	O
.	O
</s>
<s>
A	O
feature	B-General_Concept
selection	I-General_Concept
algorithm	O
can	O
be	O
seen	O
as	O
the	O
combination	O
of	O
a	O
search	O
technique	O
for	O
proposing	O
new	O
feature	O
subsets	O
,	O
along	O
with	O
an	O
evaluation	O
measure	O
which	O
scores	O
the	O
different	O
feature	O
subsets	O
.	O
</s>
<s>
The	O
simplest	O
algorithm	O
is	O
to	O
test	O
each	O
possible	O
subset	O
of	O
features	B-Algorithm
finding	O
the	O
one	O
which	O
minimizes	O
the	O
error	O
rate	O
.	O
</s>
<s>
The	O
choice	O
of	O
evaluation	O
metric	O
heavily	O
influences	O
the	O
algorithm	O
,	O
and	O
it	O
is	O
these	O
evaluation	O
metrics	O
which	O
distinguish	O
between	O
the	O
three	O
main	O
categories	O
of	O
feature	B-General_Concept
selection	I-General_Concept
algorithms	O
:	O
wrappers	O
,	O
filters	O
and	O
embedded	O
methods	O
.	O
</s>
<s>
Common	O
measures	O
include	O
the	O
mutual	O
information	O
,	O
the	O
pointwise	O
mutual	O
information	O
,	O
Pearson	O
product-moment	O
correlation	O
coefficient	O
,	O
Relief-based	B-General_Concept
algorithms	I-General_Concept
,	O
and	O
inter/intra	O
class	O
distance	O
or	O
the	O
scores	O
of	O
significance	O
tests	O
for	O
each	O
class/feature	O
combinations	O
.	O
</s>
<s>
However	O
the	O
feature	O
set	O
does	O
n't	O
contain	O
the	O
assumptions	O
of	O
a	O
prediction	O
model	O
,	O
and	O
so	O
is	O
more	O
useful	O
for	O
exposing	O
the	O
relationships	O
between	O
the	O
features	B-Algorithm
.	O
</s>
<s>
Many	O
filters	O
provide	O
a	O
feature	O
ranking	O
rather	O
than	O
an	O
explicit	O
best	O
feature	O
subset	O
,	O
and	O
the	O
cut	O
off	O
point	O
in	O
the	O
ranking	O
is	O
chosen	O
via	O
cross-validation	B-Application
.	O
</s>
<s>
One	O
other	O
popular	O
approach	O
is	O
the	O
Recursive	O
Feature	O
Elimination	O
algorithm	O
,	O
commonly	O
used	O
with	O
Support	B-Algorithm
Vector	I-Algorithm
Machines	I-Algorithm
to	O
repeatedly	O
construct	O
a	O
model	O
and	O
remove	O
features	B-Algorithm
with	O
low	O
weights	O
.	O
</s>
<s>
Embedded	O
methods	O
are	O
a	O
catch-all	O
group	O
of	O
techniques	O
which	O
perform	O
feature	B-General_Concept
selection	I-General_Concept
as	O
part	O
of	O
the	O
model	O
construction	O
process	O
.	O
</s>
<s>
The	O
exemplar	O
of	O
this	O
approach	O
is	O
the	O
LASSO	B-Algorithm
method	O
for	O
constructing	O
a	O
linear	O
model	O
,	O
which	O
penalizes	O
the	O
regression	O
coefficients	O
with	O
an	O
L1	O
penalty	O
,	O
shrinking	O
many	O
of	O
them	O
to	O
zero	O
.	O
</s>
<s>
Any	O
features	B-Algorithm
which	O
have	O
non-zero	O
regression	O
coefficients	O
are	O
'	O
selected	O
 '	O
by	O
the	O
LASSO	B-Algorithm
algorithm	O
.	O
</s>
<s>
Improvements	O
to	O
the	O
LASSO	B-Algorithm
include	O
Bolasso	O
which	O
bootstraps	O
samples	O
;	O
Elastic	O
net	O
regularization	O
,	O
which	O
combines	O
the	O
L1	O
penalty	O
of	O
LASSO	B-Algorithm
with	O
the	O
L2	O
penalty	O
of	O
ridge	O
regression	O
;	O
and	O
FeaLect	O
which	O
scores	O
all	O
the	O
features	B-Algorithm
based	O
on	O
combinatorial	O
analysis	O
of	O
regression	O
coefficients	O
.	O
</s>
<s>
AEFS	O
further	O
extends	O
LASSO	B-Algorithm
to	O
nonlinear	O
scenario	O
with	O
autoencoders	B-Algorithm
.	O
</s>
<s>
In	O
traditional	O
regression	O
analysis	O
,	O
the	O
most	O
popular	O
form	O
of	O
feature	B-General_Concept
selection	I-General_Concept
is	O
stepwise	O
regression	O
,	O
which	O
is	O
a	O
wrapper	O
technique	O
.	O
</s>
<s>
It	O
is	O
a	O
greedy	B-Algorithm
algorithm	I-Algorithm
that	O
adds	O
the	O
best	O
feature	O
(	O
or	O
deletes	O
the	O
worst	O
feature	O
)	O
at	O
each	O
round	O
.	O
</s>
<s>
In	O
machine	O
learning	O
,	O
this	O
is	O
typically	O
done	O
by	O
cross-validation	B-Application
.	O
</s>
<s>
More	O
robust	O
methods	O
have	O
been	O
explored	O
,	O
such	O
as	O
branch	B-Algorithm
and	I-Algorithm
bound	I-Algorithm
and	O
piecewise	O
linear	O
network	O
.	O
</s>
<s>
Subset	O
selection	O
evaluates	O
a	O
subset	O
of	O
features	B-Algorithm
as	O
a	O
group	O
for	O
suitability	O
.	O
</s>
<s>
Wrappers	O
use	O
a	O
search	B-Application
algorithm	I-Application
to	O
search	O
through	O
the	O
space	O
of	O
possible	O
features	B-Algorithm
and	O
evaluate	O
each	O
subset	O
by	O
running	O
a	O
model	O
on	O
the	O
subset	O
.	O
</s>
<s>
Many	O
popular	O
search	O
approaches	O
use	O
greedy	B-Algorithm
hill	B-Algorithm
climbing	I-Algorithm
,	O
which	O
iteratively	O
evaluates	O
a	O
candidate	O
subset	O
of	O
features	B-Algorithm
,	O
then	O
modifies	O
the	O
subset	O
and	O
evaluates	O
if	O
the	O
new	O
subset	O
is	O
an	O
improvement	O
over	O
the	O
old	O
.	O
</s>
<s>
Evaluation	O
of	O
the	O
subsets	O
requires	O
a	O
scoring	O
metric	O
that	O
grades	O
a	O
subset	O
of	O
features	B-Algorithm
.	O
</s>
<s>
Exhaustive	O
search	O
is	O
generally	O
impractical	O
,	O
so	O
at	O
some	O
implementor	O
(	O
or	O
operator	O
)	O
defined	O
stopping	O
point	O
,	O
the	O
subset	O
of	O
features	B-Algorithm
with	O
the	O
highest	O
score	O
discovered	O
up	O
to	O
that	O
point	O
is	O
selected	O
as	O
the	O
satisfactory	O
feature	O
subset	O
.	O
</s>
<s>
Alternative	O
search-based	O
techniques	O
are	O
based	O
on	O
targeted	B-Application
projection	I-Application
pursuit	I-Application
which	O
finds	O
low-dimensional	O
projections	O
of	O
the	O
data	O
that	O
score	O
highly	O
:	O
the	O
features	B-Algorithm
that	O
have	O
the	O
largest	O
projections	O
in	O
the	O
lower-dimensional	O
space	O
are	O
then	O
selected	O
.	O
</s>
<s>
These	O
scores	O
are	O
computed	O
between	O
a	O
candidate	O
feature	O
(	O
or	O
set	O
of	O
features	B-Algorithm
)	O
and	O
the	O
desired	O
output	O
category	O
.	O
</s>
<s>
The	O
choice	O
of	O
optimality	O
criteria	O
is	O
difficult	O
as	O
there	O
are	O
multiple	O
objectives	O
in	O
a	O
feature	B-General_Concept
selection	I-General_Concept
task	O
.	O
</s>
<s>
Many	O
common	O
criteria	O
incorporate	O
a	O
measure	O
of	O
accuracy	O
,	O
penalised	O
by	O
the	O
number	O
of	O
features	B-Algorithm
selected	O
.	O
</s>
<s>
Other	O
criteria	O
are	O
Bayesian	B-General_Concept
information	I-General_Concept
criterion	I-General_Concept
(	O
BIC	B-General_Concept
)	O
,	O
which	O
uses	O
a	O
penalty	O
of	O
for	O
each	O
added	O
feature	O
,	O
minimum	O
description	O
length	O
(	O
MDL	O
)	O
which	O
asymptotically	O
uses	O
,	O
Bonferroni	B-General_Concept
/	O
RIC	O
which	O
use	O
,	O
maximum	O
dependency	O
feature	B-General_Concept
selection	I-General_Concept
,	O
and	O
a	O
variety	O
of	O
new	O
criteria	O
that	O
are	O
motivated	O
by	O
false	B-General_Concept
discovery	I-General_Concept
rate	I-General_Concept
(	O
FDR	O
)	O
,	O
which	O
use	O
something	O
close	O
to	O
.	O
</s>
<s>
A	O
maximum	O
entropy	O
rate	O
criterion	O
may	O
also	O
be	O
used	O
to	O
select	O
the	O
most	O
relevant	O
subset	O
of	O
features	B-Algorithm
.	O
</s>
<s>
Filter	O
feature	B-General_Concept
selection	I-General_Concept
is	O
a	O
specific	O
case	O
of	O
a	O
more	O
general	O
paradigm	O
called	O
structure	B-General_Concept
learning	I-General_Concept
.	O
</s>
<s>
Feature	B-General_Concept
selection	I-General_Concept
finds	O
the	O
relevant	O
feature	O
set	O
for	O
a	O
specific	O
target	O
variable	O
whereas	O
structure	B-General_Concept
learning	I-General_Concept
finds	O
the	O
relationships	O
between	O
all	O
the	O
variables	O
,	O
usually	O
by	O
expressing	O
these	O
relationships	O
as	O
a	O
graph	O
.	O
</s>
<s>
The	O
most	O
common	O
structure	B-General_Concept
learning	I-General_Concept
algorithms	O
assume	O
the	O
data	O
is	O
generated	O
by	O
a	O
Bayesian	O
Network	O
,	O
and	O
so	O
the	O
structure	O
is	O
a	O
directed	O
graphical	O
model	O
.	O
</s>
<s>
The	O
optimal	O
solution	O
to	O
the	O
filter	O
feature	B-General_Concept
selection	I-General_Concept
problem	I-General_Concept
is	O
the	O
Markov	O
blanket	O
of	O
the	O
target	O
node	O
,	O
and	O
in	O
a	O
Bayesian	O
Network	O
,	O
there	O
is	O
a	O
unique	O
Markov	O
Blanket	O
for	O
each	O
node	O
.	O
</s>
<s>
There	O
are	O
different	O
Feature	B-General_Concept
Selection	I-General_Concept
mechanisms	O
around	O
that	O
utilize	O
mutual	O
information	O
for	O
scoring	O
the	O
different	O
features	B-Algorithm
.	O
</s>
<s>
However	O
,	O
there	O
are	O
different	O
approaches	O
,	O
that	O
try	O
to	O
reduce	O
the	O
redundancy	O
between	O
features	B-Algorithm
.	O
</s>
<s>
proposed	O
a	O
feature	B-General_Concept
selection	I-General_Concept
method	O
that	O
can	O
use	O
either	O
mutual	O
information	O
,	O
correlation	O
,	O
or	O
distance/similarity	O
scores	O
to	O
select	O
features	B-Algorithm
.	O
</s>
<s>
The	O
aim	O
is	O
to	O
penalise	O
a	O
feature	O
's	O
relevancy	O
by	O
its	O
redundancy	O
in	O
the	O
presence	O
of	O
the	O
other	O
selected	O
features	B-Algorithm
.	O
</s>
<s>
The	O
redundancy	O
of	O
all	O
features	B-Algorithm
in	O
the	O
set	O
is	O
the	O
average	O
value	O
of	O
all	O
mutual	O
information	O
values	O
between	O
the	O
feature	O
and	O
the	O
feature	O
:	O
</s>
<s>
Suppose	O
that	O
there	O
are	O
full-set	O
features	B-Algorithm
.	O
</s>
<s>
The	O
mRMR	O
algorithm	O
is	O
an	O
approximation	O
of	O
the	O
theoretically	O
optimal	O
maximum-dependency	O
feature	B-General_Concept
selection	I-General_Concept
algorithm	O
that	O
maximizes	O
the	O
mutual	O
information	O
between	O
the	O
joint	O
distribution	O
of	O
the	O
selected	O
features	B-Algorithm
and	O
the	O
classification	O
variable	O
.	O
</s>
<s>
In	O
certain	O
situations	O
the	O
algorithm	O
may	O
underestimate	O
the	O
usefulness	O
of	O
features	B-Algorithm
as	O
it	O
has	O
no	O
way	O
to	O
measure	O
interactions	O
between	O
features	B-Algorithm
which	O
can	O
increase	O
relevancy	O
.	O
</s>
<s>
This	O
can	O
lead	O
to	O
poor	O
performance	O
when	O
the	O
features	B-Algorithm
are	O
individually	O
useless	O
,	O
but	O
are	O
useful	O
when	O
combined	O
(	O
a	O
pathological	O
case	O
is	O
found	O
when	O
the	O
class	O
is	O
a	O
parity	O
function	O
of	O
the	O
features	B-Algorithm
)	O
.	O
</s>
<s>
mRMR	O
is	O
a	O
typical	O
example	O
of	O
an	O
incremental	O
greedy	B-Algorithm
strategy	O
for	O
feature	B-General_Concept
selection	I-General_Concept
:	O
once	O
a	O
feature	O
has	O
been	O
selected	O
,	O
it	O
cannot	O
be	O
deselected	O
at	O
a	O
later	O
stage	O
.	O
</s>
<s>
While	O
mRMR	O
could	O
be	O
optimized	O
using	O
floating	O
search	O
to	O
reduce	O
some	O
features	B-Algorithm
,	O
it	O
might	O
also	O
be	O
reformulated	O
as	O
a	O
global	O
quadratic	B-Algorithm
programming	I-Algorithm
optimization	O
problem	O
as	O
follows	O
:	O
</s>
<s>
where	O
is	O
the	O
vector	O
of	O
feature	O
relevancy	O
assuming	O
there	O
are	O
features	B-Algorithm
in	O
total	O
,	O
is	O
the	O
matrix	O
of	O
feature	O
pairwise	O
redundancy	O
,	O
and	O
represents	O
relative	O
feature	O
weights	O
.	O
</s>
<s>
QPFS	O
is	O
solved	O
via	O
quadratic	B-Algorithm
programming	I-Algorithm
.	O
</s>
<s>
It	O
is	O
recently	O
shown	O
that	O
QFPS	O
is	O
biased	O
towards	O
features	B-Algorithm
with	O
smaller	O
entropy	O
,	O
due	O
to	O
its	O
placement	O
of	O
the	O
feature	O
self	O
redundancy	O
term	O
on	O
the	O
diagonal	O
of	O
.	O
</s>
<s>
recommended	O
the	O
joint	O
mutual	O
information	O
as	O
a	O
good	O
score	O
for	O
feature	B-General_Concept
selection	I-General_Concept
.	O
</s>
<s>
The	O
score	O
tries	O
to	O
find	O
the	O
feature	O
,	O
that	O
adds	O
the	O
most	O
new	O
information	O
to	O
the	O
already	O
selected	O
features	B-Algorithm
,	O
in	O
order	O
to	O
avoid	O
redundancy	O
.	O
</s>
<s>
The	O
score	O
uses	O
the	O
conditional	O
mutual	O
information	O
and	O
the	O
mutual	O
information	O
to	O
estimate	O
the	O
redundancy	O
between	O
the	O
already	O
selected	O
features	B-Algorithm
(	O
)	O
and	O
the	O
feature	O
under	O
investigation	O
(	O
)	O
.	O
</s>
<s>
For	O
high-dimensional	O
and	O
small	O
sample	O
data	O
(	O
e.g.	O
,	O
dimensionality	O
>	O
and	O
the	O
number	O
of	O
samples	O
<	O
)	O
,	O
the	O
Hilbert-Schmidt	O
Independence	O
Criterion	O
Lasso	B-Algorithm
(	O
HSIC	O
Lasso	B-Algorithm
)	O
is	O
useful	O
.	O
</s>
<s>
where	O
is	O
a	O
kernel-based	O
independence	O
measure	O
called	O
the	O
(	O
empirical	O
)	O
Hilbert-Schmidt	O
independence	O
criterion	O
(	O
HSIC	O
)	O
,	O
denotes	O
the	O
trace	O
,	O
is	O
the	O
regularization	O
parameter	O
,	O
and	O
are	O
input	O
and	O
output	O
centered	O
Gram	B-Algorithm
matrices	I-Algorithm
,	O
and	O
are	O
Gram	B-Algorithm
matrices	I-Algorithm
,	O
and	O
are	O
kernel	O
functions	O
,	O
is	O
the	O
centering	O
matrix	O
,	O
is	O
the	O
-dimensional	O
identity	B-Algorithm
matrix	I-Algorithm
(	O
:	O
the	O
number	O
of	O
samples	O
)	O
,	O
is	O
the	O
-dimensional	O
vector	O
with	O
all	O
ones	O
,	O
and	O
is	O
the	O
-norm	O
.	O
</s>
<s>
The	O
optimization	O
problem	O
is	O
a	O
Lasso	B-Algorithm
problem	O
,	O
and	O
thus	O
it	O
can	O
be	O
efficiently	O
solved	O
with	O
a	O
state-of-the-art	O
Lasso	B-Algorithm
solver	O
such	O
as	O
the	O
dual	O
augmented	B-Algorithm
Lagrangian	I-Algorithm
method	I-Algorithm
.	O
</s>
<s>
The	O
correlation	O
feature	B-General_Concept
selection	I-General_Concept
(	O
CFS	O
)	O
measure	O
evaluates	O
subsets	O
of	O
features	B-Algorithm
on	O
the	O
basis	O
of	O
the	O
following	O
hypothesis	O
:	O
"	O
Good	O
feature	O
subsets	O
contain	O
features	B-Algorithm
highly	O
correlated	O
with	O
the	O
classification	O
,	O
yet	O
uncorrelated	O
to	O
each	O
other	O
"	O
.	O
</s>
<s>
The	O
following	O
equation	O
gives	O
the	O
merit	O
of	O
a	O
feature	O
subset	O
S	O
consisting	O
of	O
k	O
features	B-Algorithm
:	O
</s>
<s>
The	O
and	O
variables	O
are	O
referred	O
to	O
as	O
correlations	O
,	O
but	O
are	O
not	O
necessarily	O
Pearson	O
's	O
correlation	O
coefficient	O
or	O
Spearman	B-General_Concept
's	I-General_Concept
ρ	I-General_Concept
.	O
</s>
<s>
Hall	O
's	O
dissertation	O
uses	O
neither	O
of	O
these	O
,	O
but	O
uses	O
three	O
different	O
measures	O
of	O
relatedness	O
,	O
minimum	O
description	O
length	O
(	O
MDL	O
)	O
,	O
symmetrical	O
uncertainty	O
,	O
and	O
relief	B-General_Concept
.	O
</s>
<s>
The	O
combinatorial	O
problems	O
above	O
are	O
,	O
in	O
fact	O
,	O
mixed	O
0	O
–	O
1	O
linear	B-Algorithm
programming	I-Algorithm
problems	I-Algorithm
that	O
can	O
be	O
solved	O
by	O
using	O
branch-and-bound	B-Algorithm
algorithms	I-Algorithm
.	O
</s>
<s>
The	O
features	B-Algorithm
from	O
a	O
decision	B-Algorithm
tree	I-Algorithm
or	O
a	O
tree	O
ensemble	B-Algorithm
are	O
shown	O
to	O
be	O
redundant	O
.	O
</s>
<s>
A	O
recent	O
method	O
called	O
regularized	O
tree	O
can	O
be	O
used	O
for	O
feature	B-General_Concept
subset	I-General_Concept
selection	I-General_Concept
.	O
</s>
<s>
Regularized	O
trees	O
only	O
need	O
build	O
one	O
tree	O
model	O
(	O
or	O
one	O
tree	O
ensemble	B-Algorithm
model	O
)	O
and	O
thus	O
are	O
computationally	O
efficient	O
.	O
</s>
<s>
Regularized	O
trees	O
naturally	O
handle	O
numerical	O
and	O
categorical	O
features	B-Algorithm
,	O
interactions	O
and	O
nonlinearities	O
.	O
</s>
<s>
They	O
are	O
invariant	O
to	O
attribute	O
scales	O
(	O
units	O
)	O
and	O
insensitive	O
to	O
outliers	O
,	O
and	O
thus	O
,	O
require	O
little	O
data	B-General_Concept
preprocessing	I-General_Concept
such	O
as	O
normalization	O
.	O
</s>
<s>
A	O
metaheuristic	B-Algorithm
is	O
a	O
general	O
description	O
of	O
an	O
algorithm	O
dedicated	O
to	O
solve	O
difficult	O
(	O
typically	O
NP-hard	O
problem	O
)	O
optimization	O
problems	O
for	O
which	O
there	O
is	O
no	O
classical	O
solving	O
methods	O
.	O
</s>
<s>
Generally	O
,	O
a	O
metaheuristic	B-Algorithm
is	O
a	O
stochastic	O
algorithm	O
tending	O
to	O
reach	O
a	O
global	O
optimum	O
.	O
</s>
<s>
There	O
are	O
many	O
metaheuristics	B-Algorithm
,	O
from	O
a	O
simple	O
local	O
search	O
to	O
a	O
complex	O
global	O
search	B-Application
algorithm	I-Application
.	O
</s>
<s>
The	O
feature	B-General_Concept
selection	I-General_Concept
methods	O
are	O
typically	O
presented	O
in	O
three	O
classes	O
based	O
on	O
how	O
they	O
combine	O
the	O
selection	O
algorithm	O
and	O
the	O
model	O
building	O
.	O
</s>
<s>
They	O
are	O
based	O
only	O
on	O
general	O
features	B-Algorithm
like	O
the	O
correlation	O
with	O
the	O
variable	O
to	O
predict	O
.	O
</s>
<s>
However	O
,	O
more	O
elaborate	O
features	B-Algorithm
try	O
to	O
minimize	O
this	O
problem	O
by	O
removing	O
variables	O
highly	O
correlated	O
to	O
each	O
other	O
,	O
such	O
as	O
the	O
Fast	O
Correlation	O
Based	O
Filter	O
(	O
FCBF	O
)	O
algorithm	O
.	O
</s>
<s>
A	O
learning	O
algorithm	O
takes	O
advantage	O
of	O
its	O
own	O
variable	B-General_Concept
selection	I-General_Concept
process	O
and	O
performs	O
feature	B-General_Concept
selection	I-General_Concept
and	O
classification	O
simultaneously	O
,	O
such	O
as	O
the	O
FRMT	O
algorithm	O
.	O
</s>
<s>
This	O
is	O
a	O
survey	O
of	O
the	O
application	O
of	O
feature	B-General_Concept
selection	I-General_Concept
metaheuristics	B-Algorithm
lately	O
used	O
in	O
the	O
literature	O
.	O
</s>
<s>
Application	O
Algorithm	O
Approach	O
Classifier	O
Evaluation	B-General_Concept
Function	I-General_Concept
Reference	O
SNPs	O
Feature	B-General_Concept
Selection	I-General_Concept
using	O
Feature	O
Similarity	O
Filter	O
r2	O
Phuong	O
2005	O
SNPs	O
Genetic	B-Algorithm
algorithm	I-Algorithm
Wrapper	O
Decision	B-Algorithm
Tree	I-Algorithm
Classification	O
accuracy	O
(	O
10-fold	O
)	O
Shah	O
2004	O
SNPs	O
Hill	B-Algorithm
climbing	I-Algorithm
Filter	O
+	O
Wrapper	O
Naive	B-General_Concept
Bayesian	I-General_Concept
Predicted	O
residual	O
sum	O
of	O
squares	O
Long	O
2007	O
SNPs	O
Simulated	B-Algorithm
annealing	I-Algorithm
Naive	B-General_Concept
bayesian	I-General_Concept
Classification	O
accuracy	O
(	O
5-fold	O
)	O
Ustunkar	O
2011	O
Segments	O
parole	O
Ant	B-Algorithm
colony	I-Algorithm
Wrapper	O
Artificial	B-Architecture
Neural	I-Architecture
Network	I-Architecture
MSE	B-Algorithm
Al-ani	O
2005	O
Marketing	O
Simulated	B-Algorithm
annealing	I-Algorithm
Wrapper	O
Regression	O
AIC	O
,	O
r2	O
Meiri	O
2006	O
Economics	O
Simulated	B-Algorithm
annealing	I-Algorithm
,	O
genetic	B-Algorithm
algorithm	I-Algorithm
Wrapper	O
Regression	O
BIC	B-General_Concept
Kapetanios	O
2007	O
Spectral	O
Mass	O
Genetic	B-Algorithm
algorithm	I-Algorithm
Wrapper	O
Multiple	O
Linear	O
Regression	O
,	O
Partial	B-Algorithm
Least	I-Algorithm
Squares	I-Algorithm
root-mean-square	B-General_Concept
error	I-General_Concept
of	O
prediction	O
Broadhurst	O
et	O
al	O
.	O
</s>
<s>
1997	O
Spam	O
Binary	O
PSO	B-Algorithm
+	O
Mutation	O
Wrapper	O
Decision	B-Algorithm
tree	I-Algorithm
weighted	O
cost	O
Zhang	O
2014	O
Microarray	O
Tabu	B-Algorithm
search	I-Algorithm
+	O
PSO	B-Algorithm
Wrapper	O
Support	B-Algorithm
Vector	I-Algorithm
Machine	I-Algorithm
,	O
K	B-General_Concept
Nearest	I-General_Concept
Neighbors	I-General_Concept
Euclidean	O
Distance	O
Chuang	O
2009	O
Microarray	O
PSO	B-Algorithm
+	O
Genetic	B-Algorithm
algorithm	I-Algorithm
Wrapper	O
Support	B-Algorithm
Vector	I-Algorithm
Machine	I-Algorithm
Classification	O
accuracy	O
(	O
10-fold	O
)	O
Alba	O
2007E	O
.	O
</s>
<s>
Gene	O
Selection	O
in	O
Cancer	O
Classification	O
using	O
PSO-SVM	O
and	O
GA-SVM	O
Hybrid	O
Algorithms	O
.	O
</s>
<s>
Congress	O
on	O
Evolutionary	O
Computation	O
,	O
Singapor	O
:	O
Singapore	O
(	O
2007	O
)	O
,	O
2007	O
Microarray	O
Genetic	B-Algorithm
algorithm	I-Algorithm
+	O
Iterated	B-Algorithm
Local	I-Algorithm
Search	I-Algorithm
Embedded	O
Support	B-Algorithm
Vector	I-Algorithm
Machine	I-Algorithm
Classification	O
accuracy	O
(	O
10-fold	O
)	O
Duval	O
2009B	O
.	O
</s>
<s>
A	O
memetic	B-Algorithm
algorithm	I-Algorithm
for	O
gene	O
selection	O
and	O
molecular	O
classification	O
of	O
an	O
cancer	O
.	O
</s>
<s>
Microarray	O
Iterated	B-Algorithm
local	I-Algorithm
search	I-Algorithm
Wrapper	O
Regression	O
Posterior	O
Probability	O
Hans	O
2007C	O
.	O
</s>
<s>
Some	O
learning	O
algorithms	O
perform	O
feature	B-General_Concept
selection	I-General_Concept
as	O
part	O
of	O
their	O
overall	O
operation	O
.	O
</s>
<s>
Local	O
learning	O
based	O
feature	B-General_Concept
selection	I-General_Concept
.	O
</s>
<s>
Numeric	O
experiments	O
showed	O
that	O
the	O
method	O
can	O
achieve	O
a	O
close-to-optimal	O
solution	O
even	O
when	O
data	O
contains	O
>1M	O
irrelevant	O
features	B-Algorithm
.	O
</s>
<s>
Recommender	O
system	O
based	O
on	O
feature	B-General_Concept
selection	I-General_Concept
.	O
</s>
<s>
The	O
feature	B-General_Concept
selection	I-General_Concept
methods	O
are	O
introduced	O
into	O
recommender	O
system	O
research	O
.	O
</s>
