Database Reference
In-Depth Information
Data Set
AFS
AIS
AET
ANG
AME
ASD
17,877.3
23,527.7
44.9
398.6
5978.2
12.6
ALARM-
1000
(18533.5)
CCGA
(38.5)
(885.6)
(1.1)
(290.7)
(110.1)
(2.3)
17
,
99
0.
5
3
0,
831
.0
1
00
3
.
9
43
0
1
.
2
22
,
133
.
8
19
.
4
MDLEP
(73
.
1)
(795
.
6)
(7
0.
8)
(654
.
3)
(619
.
3)
(4
.
2)
33,836.5
45,720.0
68.1
384.4
8710.8
8.2
ALARM-
2000
(34287.9)
CCGA
(92.8)
(1,750.3)
(1.7)
(265.6)
(139.9)
(1.2)
33,932.6
56,896.6
1307.8
4046.6
25,905.8
12.9
MDLEP
(215.8)
(1,259.5)
(125.1)
(634.1)
(911.3)
(4.9)
81,033.1
111,738.0
114.1
422.1
9118.1
6.1
ALARM-
5000
(81233.4)
CCGA
(64.4)
(4,389.9)
(1.5)
(264.9)
(139.5)
(0.5)
81,287.6
134,487.2
1843.2
3946.3
29,570.8
10.7
MDLEP
(419.9)
(1,836.0)
(359.0)
(651.2)
(1,016.3)
(4.9)
158,432.4
224,246.3
204.6
422.8
10,531.5
3.4
ALARM-
10000
(158497.0)
CCGA
(16.3)
(7,070.0)
(3.6)
(286.8)
(96.2)
(0.7)
158
,
7
0
4
.
4
256
,
946
.
2
2435
.
1
3596
.
7
32
,
16
0.
8
8
.
7
MDLEP
(513.1)
(3,843.7)
(350.1)
(720.0)
(1,538.0)
(5.1)
138,854.3
217,386.1
269.8
462.7
13,635.7
11.9
ALARM-
O
(138455.0)
CCGA
(564.1)
(12,898.4)
(32.5)
(247.1)
(186.4)
(5.0)
138,913.4
252,818.4
4,09.9
4523.8
34,309.5
17.5
MDLEP
(460.8)
(5,862.0)
(2,021.3)
(482.1)
(1,327.5)
(6.9)
3413.4
3650.7
2.9
5.9
42.1
3.7
ASIA-
1000
(3416.9)
CCGA
(0.0)
(104.1)
(0.1)
(5.3)
(0.5)
(0.5)
3398.6
3590.2
76.3
79.6
656.8
3.5
MDLEP
(0.0)
(48.5)
(0.4)
(30.2)
(9.2)
(0.5)
106,541.6
114,967.2
66.4
10.1
2552.1
0.0
PRINTD-
5000
(106541.6)
CCGA
(0.0)
(900.3)
(7.5)
(3.9)
(43.8)
(0.0)
106,541.6
116,089.6
704.5
512.1
17,688.4
0.0
MDLEP
(
0.0
)
(546
.
4)
(13
.
8)
(95
.
8)
(373
.
7)
(
0.0
)
Table 6.2.
Performance comparison between CCGA and MDLEP.
optimal solution. In another setting, when we set the cutoff value to 1 (i.e.,
ignore the test results), we find that the performance of CCGA equals that
of MDLEP.
Regarding the structural difference measure (i.e., ASD), we observe that
CCGA consistently performs better than MDLEP. There are two possibil-
ities that can account for this observation: One directly relates to the CI
tests, another relates indirectly. On the one hand, the CI test phase possibly
helps to focus the search on dealing with only the correct edges (that appear
in the original structure) so that the result returned will be similar to the
original one. Although MDLEP may be successful in finding structures with
low scores, such structures may contain some wrong edges. Hence, it will be
the merit of the entire hybrid learning framework, which helps to recover
structures that closely resemble the original one. On the other hand, the ob-
servation could also be explained by the fact that better results are obtained
because searching is e
cient. In this regard, we assume that the metric di-
rectly relates to the structural difference measure. Hence, networks with good
scores will also resemble the original network. Because the searching is made
e
cient as a consequence of the reduction of search space by CI tests, we can
often find these good solutions that make ASD small.
Search WWH ::
Custom Search