Feature Selection Models for Data Classification: Wrapper Model vs Filter Model

Abstract

Data Classification is a managed learning process, which sort and organize the data into various categories to be used in more effective and efficient way. Feature selection (FS) is a noteworthy theme for the advancement of data classification process. It is regularly a basic information handling step preceding applying a learning algorithm, where the choosing of the most fitting subset of features that portrays a given classification task and the removing of irrelevant and repetitive information is occurred, hence this increases the performance of machine learning algorithms. Usually, two public methods of feature selection are used: a wrapper method in which the proposed learning algorithm itself is used to estimate how much the features are helpfully and a filter method, which assesses features heuristically, depend on the data general characteristics. In this work, the goal is to make a performance and effectiveness analysis of feature selection methodologies. This analysis handles a comparison between wrapper and filter method, by applying them on seven datasets prior to using these datasets in three classification algorithms. The results explains a set of important issues related to FS methods such as the selection taken time and the accuracy of the classification algorithms when apply FS methods to their testing datasets. The experiments were tested on seven standard datasets downloaded from multiple resources such as Machine Learning Repository and Weka tool. The experimental results have revealed that different feature selection methods can effectively enhance classification problems, but not all the time. No single method is the best for all datasets.