An Empirical Study of Wrappers for Feature Subset Selection based on a Parallel Genetic Algorithm: The Multi-Wrapper Model
AdvisorsBajic, Vladimir B.
Embargo End Date2013-09-30
Permanent link to this recordhttp://hdl.handle.net/10754/244576
MetadataShow full item record
Access RestrictionsAt the time of archiving, the student author of this thesis opted to temporarily restrict access to it. The full text of this thesis became available to the public after the expiration of the embargo on 2013-09-30.
AbstractFeature selection is the first task of any learning approach that is applied in major fields of biomedical, bioinformatics, robotics, natural language processing and social networking. In feature subset selection problem, a search methodology with a proper criterion seeks to find the best subset of features describing data (relevance) and achieving better performance (optimality). Wrapper approaches are feature selection methods which are wrapped around a classification algorithm and use a performance measure to select the best subset of features. We analyze the proper design of the objective function for the wrapper approach and highlight an objective based on several classification algorithms. We compare the wrapper approaches to different feature selection methods based on distance and information based criteria. Significant improvement in performance, computational time, and selection of minimally sized feature subsets is achieved by combining different objectives for the wrapper model. In addition, considering various classification methods in the feature selection process could lead to a global solution of desirable characteristics.
CitationSoufan, O. (2012). An Empirical Study of Wrappers for Feature Subset Selection based on a Parallel Genetic Algorithm: The Multi-Wrapper Model. KAUST Research Repository. https://doi.org/10.25781/KAUST-R9Q03