An Assessment of Nature-Inspired Algorithms for Text Feature Selection

Abstract

This paper provides a comprehensive assessment of feature selection (FS) methods that are originated from nature-inspired (NI) meta-heuristics, where two well-known filter-based FS methods are also included for comparison. The performances of the considered methods are compared on two different high-dimensional and real-world text datasets against the accuracy, the number of selected features, and computation time. This study differs from existing studies in terms of the extent of experimental analyses performed under different circumstances where classifier, feature model, and term weighting scheme are different. The results of the extensive experiments indicate that NI algorithms produce slightly different results than filter-based methods for the problem of the text FS. However, filter-based methods often provide better results by using a lower number of features and computation times.

A new modification and application of item response theory‐based feature selection for different machine learning tasks

“…One of the most popular ways to achieve dimension reduction is to apply FS in which different methods can be used to select the most informative features. Considering the evaluation criteria, the methods used in the FS step can be categorized into three types namely, filters, wrappers (e.g., nature‐inspired meta‐heuristics), and embedded 13,45 . In this paper, we use the filter‐based approach and employ the most popular 13 methods including chi‐square (CHI or $$$ {\chi}^2 $$$), mutual information (MI), and one‐way ANOVA function (AF).…”

“…From the FS perspective, CHI measures how a feature is independent of a class (i.e., label), while MI takes its maximum value if there exists a strong dependence between a feature and a class. 13,23,46 These two methods have quite a similar definition and the goodness of a feature f with respect to a class c is measured as follows: 7,13,23,46…”

“…9,12 Hence, there exist many research efforts to propose a new FS method or improve the existing one in the literature. Based on the evaluation criteria, FS methods can be grouped into three categories: 7,[9][10][11][12] filter-based, wrapper (e.g., heuristic search algorithms 13 ), and embedded methods. As the involved approach in this study, filter-based FS methods use various distance and information metrics for feature evaluation, where features are ranked based on the employed metric such as chi-square, mutual information, odds ratio, information gain, f-statistics, and so on.…”

“…7,19 Hence, the main applications of IRT can be found in educational testing in which analysts are interested in measuring examinees' ability using a test that consists of several questions. 16,18 Although there exists many research efforts 2,12,13,[20][21][22][23] on the task FS process of ML and IRT models have a wide range of applications (e.g., ability assessment, item analysis, test scoring, etc. ), the number of studies considering combining the powers of these two fields is still limited.…”

“…Although there exists many research efforts 2,12,13,20‐23 on the task FS process of ML and IRT models have a wide range of applications (e.g., ability assessment, item analysis, test scoring, etc. ), the number of studies considering combining the powers of these two fields is still limited.…”