Database Reference
In-Depth Information
to the nature of the literals it will not always be 0. If the end of the series
is before the beginning the interval, the result will be 0. If the end of the
series is in the interval, the result depends on the predicate:
•
For
increases
and
decreases
the result will be always 0, because only the
extremes of the interval are considered.
•
For
stays
the result will be
1 if the available points are not in the allowed
range. In other case the result will be 0.
−
•
For
always
, the result will be
1 if there is some available point in the
interval that is out of the region. In other case the result will be 0.
−
•
For
sometimes
, the result will be 1 if there is some available point in the
interval that is out of the region. In other case the result will be 0.
•
For
true percentage
, the result will be 1 if the series has already enough
points in the interval inside the region. In other case the result will be 0.
5. Early Classification
The capability of obtaining an initial classification as soon as the available
data give some cue about the possible class they belong to, is a very desir-
able property if the classifiers are to be used in a dynamic environment were
data are generated and analyzed on line. Now, the setting is different to
the variable length series learning problem of the previous paragraph: the
classifier has already been obtained, maybe form series of different length,
and the variations occur on the length of the series to be classified, from
partial examples to full series.
Somewhat surprisingly, this early classification ability can be obtained
without modifying the learning method, exploiting the same ideas that
allow learning from series of different length. When a partial example is
presented to the classifier some of its literals can be evaluated, because their
intervals refer to areas that are already available in the example. Certainly,
some literals cannot be evaluated because the intervals that they refer to
are still not available for the example: its value is still unknown. Given
that the classifier consists of a linear combination of literals, the simple
omission of literals with unknown value allows to obtain a classification
from the available data. The classification given to a partial example will
be the linear combination of those literals that have known results.
A very simple approach to identify literals with unknown values would
be to abstain whenever there are point values that are still unknown for
the example. Nevertheless, if the values of some points in the interval are
