Affiliations: Department of Computer Science, University of Seville,
Avda. Reina Mercedes S/n, 41012 Sevilla, Spain. E-mail: {rruiz, riquelme,
aguilar}@lsi.us.es
Abstract: The attribute selection techniques for supervised learning, used in
the preprocessing phase to emphasize the most relevant attributes, allow making
models of classification simpler and easy to understand. Depending on the
method to apply: starting point, search organization, evaluation strategy, and
the stopping criterion, there is an added cost to the classification algorithm
that we are going to use, that normally will be compensated, in greater or
smaller extent, by the attribute reduction in the classification model. The
method proposed in this work utilizes a measure based on projections to guide
the selection of the attributes. The algorithm (SOAP: Selection of Attributes
by Projection) has some interesting characteristics: lower computational cost
(O(mn log n) m attributes and n examples in the data set) with respect to other
typical algorithms due to the absence of distance and statistical calculations;
its applicability to any labelled data set, that is to say, it can contain
continuous and discrete variables, with no need for transformation. The
performance of SOAP is analysed in two ways: percentage of reduction and
classification. SOAP has been compared to CFS [4] and ReliefF [8]. The results
are generated by C4.5 before and after the application of the algorithms.
