Graphics Reference
In-Depth Information
Chapter 7
Feature Selection
Abstract
In this chapter, one of the most commonly used techniques for
dimensionality and data reduction will be described. The feature selection problem
will be discussed and the main aspects and methods will be analyzed. The chapter
starts with the topics theoretical background (Sect.
7.1
), dividing it into the major
perspectives (Sect.
7.2
) and the main aspects, including applications and the eval-
uation of feature selections methods (Sect.
7.3
). From this point on, the successive
sections make a tour from the classical approaches, to the most advanced proposals,
in Sect.
7.4
. Focusing on hybridizations, better optimization models and derivatives
methods related with feature selection, Sect.
7.5
provides a summary on related and
advanced topics, such as feature construction and feature extraction. An enumeration
of some comparative experimental studies conducted in the specialized literature is
included in Sect.
7.6
.
7.1 Overview
competence of most learning algorithms, especially due to the fact that they usually
are computationally expensive. Feature selection (FS) is an effective form of dealing
with DR.
We have to answer what is the result of FS and why we need FS. For the first
question, the effect is to have a reduced subset of features from the original set; for the
latter, the purposes can vary: (1) to improve performance (in terms of speed, predictive
power, simplicity of the model); (2) to visualize the data for model selection; (3) to
reduce dimensionality and remove noise. Combining all these issues, we can define
FS as follows [
29
]:
Definition 7.1
Feature Selection
is a process that chooses an optimal subset of fea-
tures according to a certain criterion.
The criterion determines the details of evaluating feature subsets. The selection of
the criterion must be done according to the purposes of FS. For example, an optimal
subset could be a minimal subset that could give the best estimate of predictive
accuracy.
