Authors:
Camelia Vidrighin Bratu
and
Rodica Potolea
Affiliation:
Technical University of Cluj-Napoca, Romania
Keyword(s):
Feature Selection, Classification, Baseline Accuracy, Combined Method.
Related
Ontology
Subjects/Areas/Topics:
Business Analytics
;
Communication and Software Technologies and Architectures
;
Data Engineering
;
Data Warehouses and Data Mining
;
e-Business
;
Enterprise Information Systems
Abstract:
Feature selection is an important step in any data mining process, for many reasons. In this paper we consider the improvement of the prediction accuracy as the main goal of a feature selection method. We focus on an existing
3-step formalism, including a generation procedure, evaluation function and validation procedure. The performance evaluations have yielded that no individual 3-tuple (generation, evaluation and validation procedure) can be identified such that it achieves best performance on any dataset, with any learning algorithm. Moreover, the experimental results suggest the possibility of tackling a combined approach to the feature selection problem. So far we have experienced with the combination of several generation procedures, but we believe that the evaluation functions can also be successfully combined.