Graphics Reference
In-Depth Information
•
Semi-Supervised Discretization:
A first attempt to discretize data in semi-
supervised classification problems has been devised in [
14
], showing that it is
asymptotically equivalent to the supervised approach.
The researchmentioned in this section is out of the scope of this topic.We point out
that the main objective of this chapter is to give a wide overview of the discretization
methods found in the literature and to conduct an experimental comparison of the
most relevant discretizers without considering external and advanced factors such as
those mentioned above or derived problems from classic supervised classification.
9.3 Properties and Taxonomy
This section presents a taxonomy of discretization methods and the criteria used
for building it. First, in Sect.
9.3.1
, the main characteristics which will define the
categories of the taxonomy will be outlined. Then, in Sect.
9.3.2
, we enumerate the
discretization methods proposed in the literature and we will consider by using both
their complete name and abbreviated name together with the associated reference.
Finally, we present the taxonomy.
9.3.1 Common Properties
This section provides a framework for the discussion of the discretizers presented in
the next subsection. The issues discussed include several properties involved in the
structure of the taxonomy, since they are exclusive to the operation of the discretizer.
Other, less critical issues such as parametric properties or stopping conditions will
be presented although they are not involved in the taxonomy. Finally, some criteria
will also be pointed out in order to compare discretization methods.
9.3.1.1 Main Characteristics of a Discretizer
In [
36
,
51
,
75
,
123
], various axes have been described in order to make a cate-
gorization of discretization methods. We review and explain them in this section,
emphasizing the main aspects and relations found among them and unifying the
notation. The taxonomy presented will be based on these characteristics:
•
Static versus Dynamic:
This characteristic refers to the moment and independence
which the discretizer operates in relation to the learner. A dynamic discretizer
acts when the learner is building the model, thus they can only access partial
information (local property, see later) embedded in the learner itself, yielding
compact and accurate results in conjuntion with the associated learner. Otherwise,
a static discretizer proceeds prior to the learning task and it is independent from
