Database Reference
In-Depth Information
known, it is sometimes possible to know the result of the literal. This is done
in a similar way than when evaluating literals for variable length series.
6. Experimental Validation
The characteristics of the data sets used to evaluate the behavior of the
learning systems are summarized in Table 2. If a training test partition
it specified, the results are obtained by averaging 5 experiments. In other
case, 10 fold stratified cross validation was used.
Table 3 shows the obtained results for different data sets with different
combination of predicates. The setting named
Relative
uses the predicates
increases
,
decreases
and
stays
. The setting
Interval
also includes these liter-
als, and
true percentage
. It does not include always and sometimes, because
they can be considered as special cases of
true percentage
.
Normally, 100 literals were used. For some data sets this value is
more than enough, and it shows clearly that the method does not over-
fit. Nevertheless, the results of some data set can be improved using more
Table 2.
Characteristics of the data sets.
Classes
Variables
Examples
Training/
Length
Test
Min.
Average
Max.
CBF
3
1
798
10 fold CV
128
128.00
128
CBF-Var
3
1
798
10 fold CV
64
95.75
128
Control
6
1
600
10 fold CV
60
60.00
60
Control-Var
6
1
600
10 fold CV
30
44.86
60
Trace
16
4
1600
800/800
67
82.49
99
Auslan
10
8
200
10 fold CV
33
49.34
102
Table 3.
Results (error rates) for the different data sets using the different literals.
Data set
Literals
Point
Relative
Always/
True
Interval
sometimes
percentage
CBF
100
3.51
2.28
1.38
0.63
0.50
CBF-Var
100
3.49
1.61
2.75
1.25
1.13
Control
100
4.00
1.17
1.00
1.33
0.83
Control-Var
100
13.00
7.00
5.00
4.83
5.00
Trace
100
72.00
3.00
3.83
8.70
0.78
Trace
200
70.98
0.03
2.65
6.78
0.20
Gloves
100
8.00
5.50
5.50
3.00
4.50
Gloves
200
7.50
6.50
3.50
2.50
2.50
Gloves
300
5.00
4.00
4.50
2.50
1.50
