[Paper Review] FSinR: an exhaustive package for feature selection

Francisco Aragón-Royón, Alfonso Jiménez-Vílchez+2 more|arXiv (Cornell University)|Feb 24, 2020

Machine Learning and Data Classification30 references10 citations

TL;DR

FSinR is an R package that provides a comprehensive, unified framework for feature selection by integrating 16 widely used filter methods and 238 wrapper methods from the caret package, combined with 13 search algorithms. It enables exhaustive, flexible, and user-friendly feature selection across diverse data types and machine learning tasks, offering superior robustness, extensibility, and integration compared to existing CRAN packages.

ABSTRACT

Feature Selection (FS) is a key task in Machine Learning. It consists in selecting a number of relevant variables for the model construction or data analysis. We present the R package, FSinR, which implements a variety of widely known filter and wrapper methods, as well as search algorithms. Thus, the package provides the possibility to perform the feature selection process, which consists in the combination of a guided search on the subsets of features with the filter or wrapper methods that return an evaluation measure of those subsets. In this article, we also present some examples on the usage of the package and a comparison with other packages available in R that contain methods for feature selection.

Motivation & Objective

• To address the lack of a comprehensive, extensible, and user-friendly R package that unifies diverse feature selection methods.
• To provide a robust and scalable solution for feature selection that supports both filter and wrapper methods with multiple search strategies.
• To integrate all models available in the caret package as wrapper evaluation measures, enhancing methodological diversity and reliability.
• To improve usability through a consistent, uniform interface and internal type checking for robust parameter handling.
• To offer a tool that supports various data types (numerical, categorical) and problem types (classification, regression), increasing versatility.

Proposed method

• The package implements 16 standard filter methods (e.g., correlation-based, variance threshold, information gain) for independent feature evaluation.
• It leverages all 238 classification and regression models from the caret package as wrapper evaluation measures, enabling model-driven feature assessment.
• It combines these methods with 13 search algorithms (e.g., best-first, genetic algorithms, recursive feature elimination) to explore feature subset spaces efficiently.
• The package uses a uniform function interface across all methods, ensuring consistent and intuitive usage for users.
• Internal validation checks types and compatibility of input parameters to improve reliability and reduce user error.
• The framework supports both classification and regression tasks and handles mixed data types (numerical, categorical) seamlessly.

Experimental results

Research questions

• RQ1 How can a unified R package be designed to integrate a broad range of filter and wrapper feature selection methods with diverse search strategies?
• RQ2 What advantages does integrating all caret models as wrapper measures offer in terms of methodological coverage and performance robustness?
• RQ3 How does FSinR compare in functionality, usability, and extensibility to existing CRAN packages focused on feature selection?
• RQ4 To what extent does the package’s consistent interface and internal validation improve user experience and reduce implementation errors?
• RQ5 Can a feature selection tool that combines exhaustive search with diverse evaluation measures outperform specialized or limited-functionality alternatives?

Key findings

• FSinR supports 16 filter methods and 238 wrapper methods from the caret package, significantly expanding methodological coverage compared to most existing R packages.
• The package enables the combination of any filter or wrapper method with any of 13 search algorithms, offering a highly flexible and exhaustive feature selection pipeline.
• Unlike FSelector, which lacks search integration and returns only selected features, FSinR provides detailed evaluation metrics and process transparency.
• The package’s uniform interface and internal type checking enhance usability and reduce errors, making it more robust than alternatives like Boruta or spSFR that offer only single-method solutions.
• FSinR outperforms other packages in extensibility and versatility, supporting all data types and problem types (classification, regression) with consistent, documented functionality.
• The integration of caret’s models as wrappers ensures high reliability and access to state-of-the-art evaluation techniques, making FSinR a comprehensive solution for feature selection in machine learning.