Information Technology Reference
In-Depth Information
Algorithm 1.
StARMiner Algorithm
Input:
Dataset
T
of image transactions structured as
where
x
represents
the image category and
f
i
an image feature; thresholds
Δμ
min
,
σ
max
,
γ
min
,
Output:
The mined rules
1. Scan dataset
T
2.
for
each feature
f
i
do
3.
for
each category
x
j
do
4. calculate
μ
f
i
(
T
x
j
)and
μ
f
i
(
T − T
x
j
)
5.
end for
6.
end for
7. Scan dataset
T
8.
for
each feature
f
i
do
9.
for
each category
x
j
do
10. calculate
σ
f
i
(
T
x
j
)and
sigma
f
i
(
T − T
x
j
)
11. calculate
Z
value
12.
{x
,
f
1
,
f
2
,
f
n
}
if
(
μ
f
i
(
T
x
j
)
− μ
f
i
(
T
− T
x
j
))
≥ Δμ
min
And
σ
f
i
(
T
x
j
)
≤ σ
max
And
(
Z<Z
1
or
Z>Z
2
)
then
13. write
x
j
→ f
i
,
μ
f
i
(
T
x
j
),
μ
f
i
(
T − T
x
j
),
σ
f
i
(
T
x
j
),
σ
f
i
(
T − T
x
j
)
14.
end if
15.
end for
16.
end for
(lines 7 to 16) calculates the standard deviation for each feature and the
Z
value,
used in the hypotheses test. The restrictions of interest are processed in lines
11 and 12. A rule is returned only if it satisfies the input thresholds (
Δμ
min
,
σ
max
,
γ
min
). The complexity of StARMiner is
Θ
(
ckN
), where
N
is the number
of instances of the dataset, k is the number of features, and
c
is the number
of categories.
Criterion 1
is employed to perform feature selection using the
StARMiner results.
Criterion 1.
The features in the set of rules returned by StARMiner algorithm
are selected as the most relevant ones.
The StARMiner algorithm makes possible to find rules that properly cate-
gorizes the images. That is, the algorithm spots the features with high power
on differentiating image categories, since they have a particular and uniform
behavior in images of a given category. This is important, because the features
(or attributes) that present a uniform behavior to every image in the dataset,
independently of the image category, do not contribute to categorize them, and
should be eliminated.
To validate the StARMiner algorithm we used a procedure composed of three
steps illustrated in Figure 7.3. These steps are detailed as follows.
Since StARMiner is a supervised feature selection algorithm, the image
dataset is divided in the training set and the test set. The training set is sub-
mitted to the feature selection algorithm, where the test set is used to evaluate
the method in the content-based queries.
