Graphics Reference
In-Depth Information
Temp = S
{
X
i
}
If
A
then
go
_
on
=
θ
If
S
≤
S
and
go
_
on
then
S=Temp
i=i+1
else
go
_
on
=
θ
Temp
<
- Thin-out selection phase (THIN). It thins out points that are not important to
the decision boundary of the resulting 1-NN rule. This phase aims at selecting
a small number of instances without negatively affecting the 1NN empirical
error.
S
with in-degree
G
bc
>
S
f
={
x
0
}
S
pr e
v
=
S
S
1
=
S
S
f
go
_
on
=1
While
go
_
on
do
S
t
={
a
\
S
1
with in-degree
G
S
t
bc
>
θ
and with in-degree
G
S
pr e
v
bc
or in
G
S
pr e
v
>
0
}
w
c
S
f
S
t
S
f
go
_
on
=
<
If
go
_
on
then
S
f
=
S
f
S
t
S
p
re
v
=
S
1
S
1
=
S
\
S
f
8.4.3.4 Wrapper
•
Backward Sequential Edition (BSE)
[
125
]—The BSE algorithm starts with
S=TR. Each instance
X
i
is tested to find how the performance of the KNN is
increased when
X
i
is removed from
S
. The instance in which removal causes
the best increase in the performance is finally deleted from
S
, and the process is
repeated until no further increases in the performance of the classifier are found.
To increase the efficiency of the method, the authors suggested the use of ENN or
DROP procedures as a first stage of the BSE algorithm.
8.4.3.5 Batch
•
(
)
Iterative Case Filtering (ICF)
[
15
]—ICF defines
local set L
which con-
tain all cases inside largest hypersphere centered in
x
i
such that the hypersphere
x
