Database Reference
In-Depth Information
to run the transformation any time when the
read
method is called (see previous
section).
Both types of transformations have advantages and disadvantages: in case of
static transformations, the transformation is done only once, and then the target
stream contains the complete transformed data. The disadvantage of this approach
is that we need two streams to be supported (and usually about twice of memory
amount). This means that static transformations are optimal
in speed but
non-optimal in memory consumption.
For dynamic transformations via mining filter streams, we do not need additional
memory, but any time when we access the data from the stream using the
read
method, the transformation of the current vector is carried out again. Hence,
dynamic transformations are non-optimal in time but optimal in memory. The static
and dynamic transformations represent the classic dilemma that increased speed
requires increased memory, and vice versa.
There are many vector transformations implemented into XELOPES based on an
extensive CWM framework. For general stream transformations, if they are not
based on vector transformations,
there is no further framework provided in
XELOPES.
Basis Transformations
Basis transformations are very important but also somewhat abstract. Luckily for
most applications, the XELOPES user does not have to care about basis trans-
formations because they are automatically executed internally. However, since
basis transformations are an important part of this topic, we will go more into detail.
For basis transformations, we need to transform metadata (basis) and mining
vectors (coordinates) of the application data with respect to the metadata of the
mining model, i.e., the metadata of the training data set.
The required basis transformation is addressed by the class
MetaDataOperations
which is a singleton class owned by the metadata class
MiningDataSpecification
.
Thus, each
MiningDataSpecification
object owns an object
MetaDataOperations
to
transform another
MiningDataSpecification
and appendant mining vectors into its
own basis. In addition,
MiningDataSpecification
contains methods like
equals
,
subset
, and
superset
for comparison with another
MiningDataSpecification
object.
In the same way, each
CategoricalAttribute
owns a singleton class
Categorica-
lAttributeOperations
for basis transformations (because here the categories are
referred to as basis of the categorical attribute) from another categorical attribute
into the current one and for comparisons. Moreover, even categorical attributes of
unboundedCategories
type are supported by adaptive basis transformation. This is
very important because it allows to apply basis transformations, e.g., to categorical
attributes of live mining input streams which continuously deliver new categories.
An example is the application of a mining model to a large customer database
where customers are continuously added during the application process.
