Database Reference
In-Depth Information
Table 10.
Results for the
Auslan
data set.
Error/Number of literals
Literals
30
60
90
120
150
180
210
240
270
300
Points
16.50
11.50
8.00
8.50
8.00
8.00
7.50
6.50
6.00
5.00
Intervals
11.00
7.50
4.00
4.00
3.50
3.00
3.00
2.50
2.00
1.50
Error/Series length percentage
Percentage
10
20
30
40
50
60
70
80
90
100
Points
76.00
42.00
10.00
5.50
5.00
5.00
5.00
5.00
5.00
5.00
Intervals
82.50
41.50
13.00
3.50
2.00
1.50
1.50
1.50
1.50
1.50
10 classes and 20 examples of each class. The minimum length is 33 and
the maximum is 102.
The results are shown in Figure 9 and in Table 10. The best result
reported in [Kadous (1999)] is an error of 2.50, using event extraction,
event clustering and Naıve Bayes Classifiers. Our result is 1.5 using 300
interval based literals.
With respect to the results for early classification, it is necessary to
consider two facts. First, for this problem it is not practical to use early
classification, because the examples are generated very quickly. The second
fact is that the used percentage is for the longest series, and in this data
set the lengths vary a lot. Hence, the obtained results for a length of 50%,
include many examples that are already completed.
7. Conclusions
This chapter has presented a novel approach to the induction of multivariate
time series classifiers. It summarizes a supervised learning method that
works boosting very simple base classifiers. From only one literal, the base
classifiers, boosting creates classifiers consisting of a linear combination of
literals.
The learning method is highly domain independent, because it only
makes use of very general techniques, like boosting, and only employs very
generic descriptions of the time series, interval based literals. In this work we
have resorted to two kind of interval predicates: relative and region based.
Relative predicates consider the evolution of the values in the interval, while
region based predicates consider the occurrence of the values of a variable in
a region during an interval. These kind of predicates, specifically design for
