[Paper Review] LOFS: Library of Online Streaming Feature Selection

Kui Yu, Wei Ding+1 more|arXiv (Cornell University)|Mar 2, 2016

Data Stream Mining Techniques11 references3 citations

TL;DR

LOFS is the first comprehensive open-source MATLAB library for online streaming feature selection, implementing state-of-the-art algorithms for selecting individual and grouped features as they arrive sequentially. It enables efficient, incremental feature selection in high-dimensional big data scenarios, with modular design supporting extensibility and performance evaluation via statistical tests.

ABSTRACT

As an emerging research direction, online streaming feature selection deals with sequentially added dimensions in a feature space while the number of data instances is fixed. Online streaming feature selection provides a new, complementary algorithmic methodology to enrich online feature selection, especially targets to high dimensionality in big data analytics. This paper introduces the first comprehensive open-source library for use in MATLAB that implements the state-of-the-art algorithms of online streaming feature selection. The library is designed to facilitate the development of new algorithms in this exciting research direction and make comparisons between the new methods and existing ones available.

Motivation & Objective

• To address the challenge of high-dimensional data in big data analytics by enabling online feature selection as features arrive sequentially.
• To provide a unified, open-source platform for implementing and comparing state-of-the-art online streaming feature selection algorithms.
• To support both individual feature selection and group-based feature selection where features arrive in dynamic, interrelated groups.
• To facilitate research and development of new online feature selection algorithms through a modular, extensible library architecture.
• To enable statistical comparison of algorithms using standard metrics and non-parametric tests like Friedman and Nemenyi.

Proposed method

• LOFS employs a three-module architecture: Correlation Measure (CM), Learning, and Statistical Comparison (SC) modules.
• The CM module computes feature correlations using Chi-square, G², Fisher’s Z, and mutual information tests for discrete and continuous data.
• The Learning module implements individual feature selection via Alpha-investing, OSFS, Fast-OSFS, and SAOLA; group feature selection via group-SAOLA.
• The SC module evaluates performance using metrics like accuracy, AUC, kappa, and compactness, and applies non-parametric tests for algorithm comparison.
• The library is implemented in MATLAB with standardized coding practices, enabling easy extension and integration of new algorithms.
• LOFS supports future integration of ARFF and LIBSVM file formats for broader data compatibility.

Experimental results

Research questions

• RQ1 How can online feature selection be effectively extended to handle sequentially arriving features in high-dimensional data streams?
• RQ2 What is the performance gain of online streaming feature selection over traditional batch feature selection methods in dynamic data environments?
• RQ3 How can group-wise feature dependencies be modeled and selected efficiently in an online setting?
• RQ4 Can a modular, open-source library accelerate the development and benchmarking of new online feature selection algorithms?
• RQ5 What statistical methods are most effective for comparing multiple online feature selection algorithms across diverse datasets?

Key findings

• LOFS is the first comprehensive open-source library specifically designed for online streaming feature selection, supporting both individual and grouped feature selection.
• The library integrates state-of-the-art algorithms including Alpha-investing, OSFS, Fast-OSFS, SAOLA, and group-SAOLA, all implemented in MATLAB.
• LOFS enables statistical comparison of algorithms using non-parametric tests such as Friedman and Nemenyi, enhancing reliability of performance evaluation.
• Empirical comparisons in the original work show that online streaming feature selection methods outperform traditional batch methods like FCBF, SPSF-LAR, and GDM in dynamic, high-dimensional settings.
• The modular architecture allows researchers to easily extend LOFS with new algorithms, promoting community-driven development and reproducibility.
• Future versions will support ARFF and LIBSVM file formats, increasing data interoperability and usability across machine learning platforms.