Graphics Reference
In-Depth Information
Due to the fact that most of the efforts in research are devoted to PS, we focus
the rest of this chapter on this issue. It is very arduous to give an exact number
of proposals belonging specifically to each of the two families mentioned above.
When an IS method is proposed, the first step is to improve instance-based learners.
When tackling TSS, the type of learning method is usually fixed to combine with the
instance selector. Few methods are proposed thinking in both processes, although our
experience allows us to suggest that any filter should work well for any DM model,
mainly due to the low reductions rates achieved and the efforts in removing noise
from the data with these kind of methods. An proper estimation of proposals reported
in the specialized literature specifically considered for TSS may be around 10 % of
the total number of techniques. Even though PS monopolizes almost all efforts in IS,
TSS currently shows an upward trend.
8.3 Prototype Selection Taxonomy
This section presents the taxonomy of PS methods and the criteria used for building
it. First, in Sec.
8.3.1
, the main characteristics which will define the categories of the
taxonomy will be outlined. In Sec.
8.3.2
, we briefly enumerate all the PS methods
proposed in the literature. The complete and abbreviated name will be given together
with the reference. Finally, Sec.
8.3.3
, presents the taxonomy.
8.3.1 Common Properties in Prototype Selection Methods
This section provides a framework for the discussion of the PS methods presented in
the next subsection. The issues discussed include order of the search, type of selection
and evaluation of the search. These mentioned issues are involved in the definition of
the taxonomy, since they are exclusive to the operation of the PS algorithms. Other
classifier-dependent issues such as distance functions or exemplar representation will
be presented. Finally, some criteria will also be pointed out in order to compare PS
methods.
8.3.1.1 Direction of Search
When searching for a subset
S
of prototypes to keep from training set
TR
, there are
a variety of directions in which the search can proceed:
•
Incremental:
An incremental search begins with an empty subset
S
, and adds each
instance in
TR
to
S
if it fulfills some criteria. In this case, the algorithm depends
on the order of presentation and this factor could be very important. Under such a
scheme, the order of presentation of instances in
TR
should be random because by
