Database Reference
In-Depth Information
(
)
, the class labels of the training data are used to define
the scores of individual features. Such scores can effectively reflect the feature's
capability in discriminating positive and negative classes. A feature's score can be
effectively represented by its SNR (Signal-to-Noise Ratio) score, which is defined
as the ratio of the signal (inter-class distinction) and noise (intra-class) perturbation
respectively. We adopt
In order to obtain
y
i
r
{
μ
+
,
μ
−
}
{
˃
+
,
˃
−
}
to denote respectively the class-
conditional means and standard deviation of the (positive/negative) classes. These
parameters are computed from the row vectors of
H
in Eq. (
5.23
). The Fisher
Discriminant Ratio (FDR) [
165
] is utilized for the calculation of the SNR score:
and
2
(
μ
j
−
μ
j
)
FDR
(
j
)=
(5.25)
˃
j
2
˃
j
2
+
˃
j
represent the class-conditional means and standard
deviations of the
j
-th feature, respectively.
In this case, the magnitude of scores for all features, i.e.,
μ
j
,
μ
j
,
˃
j
, and
where
M
j
1
can
be used to measure the relevancy of their corresponding features. More exactly, the
features
h
j
M
j
=
1
{|
FDR
(
j
)
|}
=
M
j
=
1
, and a
fraction of the lowest-ranked features will be eliminated. The selected feature set is
denoted as:
can be ranked accordingly to their scores
{|
FDR
(
j
)
|}
{
y
1
,...,
y
i
,
y
i
+
1
...,
y
m
},|
FDR
(
i
)
| > |
FDR
(
i
+
1
)
|
(5.26)
This feature set is further divided into two subsets:
⊨
⊬
y
1
,
y
2
,...,
y
T
1
y
T
1
+
1
,
y
T
1
+
2
,...,
y
m
(5.27)
⊩
⊭
Y
Y
where
T
1
is the threshold value, and the subset
Y
contains the features having SNR
scores greater than those of the subset
Y
.
The feature selection process can be summarized by its flowchart shown in
Fig.
5.3
. In the final step, the selected features in
Y
and
Y
are used as a weighting
factor for the original BoW features of the query,
h
q
=
h
1
,...,
h
q
M
t
. According to
the indexes of features in
Y
, and
Y
,the
j
-th element of the query is modified by:
⊧
⊨
h
j
+
ʵ
Y
1
y
j
,
y
j
∈
h
j
=
h
j
−
ʵ
2
y
j
,
Y
y
j
∈
(5.28)
⊩
h
j
otherwise
Search WWH ::
Custom Search