Databases Reference
In-Depth Information
Tabl e 1 .
Classification accuracy (
α
= 50%,
σ
=1%)
Datasets
CSA
One-by-one approach Randomised selector
k
=1
k
=10
k
=1
k
=10
k
=5
k
=50
adult.D97.N48842.C2
80.83
83.87
76.88
81.95
81.85
anneal.D73.N898.C6
91.09
89.31
91.09
90.20
91.31
auto.D137.N205.C7
61.76
64.71
59.80
64.71
58.82
breast.D20.N699.C2
89.11
87.68
89.11
90.83
92.55
connect4.D129.N67557.C3
65.83
66.78
65.87
66.34
66.05
cylBands.D124.N540.C2
65.93
69.63
63.70
67.41
67.78
flare.D39.N1389.C9
84.44
84.01
84.29
84.44
84.29
glass.D48.N214.C7
58.88
64.49
52.34
64.49
64.49
heart.D52.N303.C5
58.28
58.28
56.29
60.26
59.60
hepatitis.D56.N155.C2
68.83
68.83
66.23
75.32
72.72
horseColic.D85.N368.C2
72.83
77.72
80.43
80.43
81.52
ionosphere.D157.N351.C2
85.14
84.00
90.29
88.57
93.14
iris.D19.N150.C3
97.33
97.33
97.33
97.33
97.33
led7.D24.N3200.C10
68.38
62.94
68.38
68.89
69.94
letRecog.D106.N20000.C26
30.29
29.41
31.19
29.36
30.92
mushroom.D90.N8124.C2
99.21
98.45
98.82
99.21
98.45
nursery.D32.N12960.C5
80.35
76.85
76.17
80.20
81.11
pageBlocks.D46.N5473.C5
90.97
91.74
90.97
91.74
90.97
pima.D38.N768.C2
73.18
73.18
73.18
73.44
73.44
soybean-large.D118.N683.C19
85.92
81.23
86.51
84.46
84.75
ticTacToe.D29.N958.C2
71.61
68.48
72.03
71.19
73.28
waveform.D101.N5000.C3
61.60
58.92
55.96
59.52
57.20
wine.D68.N178.C3
53.93
83.15
71.91
83.15
85.39
zoo.D42.N101.C7
76.00
86.00
78.00
90.00
86.00
Average
73.82
75.29
74.03
76.81
76.79
(only the most significantly CAR for each class is mined) and applying the
“one-by-one” rule mining approach, the average accuracy of classification
throughout the 24 datasets is 75.29%. When substituting the value of 1 by a
value of 10 (the best ten significant CARs for each class are identified), the av-
erage accuracy, using the “one-by-one” rule mining approach, is 74.03%. Note
that the average accuracies are higher than the average accuracy of classifica-
tion obtained by the well-established CSA ordering approach, which is 73.82%.
Furthermore when dealing with the randomised selector based rule mining ap-
proach, and choosing a value of 1 as the value for
k
and a value of 5 as the
value for
k
(only the most significantly CAR for each class is mined, based on
the existence of five potential significant CARs for each class in
), the aver-
age accuracy throughout the 24 datasets can be obtained as 76.81%. Note that
in the randomised experiment process, we always run several tests (i.e., 8-10
tests) for each dataset, and catch the best result. When substituting the value
of 1 by a value of 10, and the value of 5 by a value of 50 (the best ten significant
R
