MFLRA improves multi-label text classification across ten benchmarks

A computational study found that a new feature-selection method called Multi-order Feature–Label Relevance and Ambiguity, or MFLRA, outperformed eight comparison methods across multiple evaluation metrics on 10 benchmark datasets for multi-label text classification. The study reported an average accuracy of 0.2995 while combining direct feature–label relationships, second-order associations, label uncertainty and redundancy within a unified selection procedure. The research appeared in Multimedia Tools and Applications as Volume 85, article 818, on 20 October 2026.

The authors developed MFLRA in response to limitations they identified in existing feature-selection methods for high-dimensional multi-label text data. Those approaches, according to the paper, may retain redundancy in the label space, depend exclusively on first- or second-order feature–label associations, or neglect uncertainty among labels. The proposed method unifies first- and second-order relevance so that both direct associations and interactions are represented, while an analysis of interactions among labels is used to reduce redundancy and noise.

At the center of the approach is a pseudo-importance matrix. Elastic Net regression contributes to the matrix by handling multicollinearity, entropy-based weighting models uncertainty in the labels, and feature interactions are included to improve robustness. The resulting matrix is processed with CODAS, a multi-criteria decision method that ranks and selects features. The selected features then feed into MLKNN, a nearest-neighbor classifier, with its parameters optimized through particle swarm optimization, or PSO. In this sense, the reported contribution lies in the feature-selection pipeline rather than in a newly proposed text classifier alone.

Experiments on 10 benchmark datasets compared the complete MFLRA pipeline with eight state-of-the-art methods across multiple evaluation metrics. The authors report that MFLRA consistently outperformed the comparison methods and achieved an average accuracy of 0.2995. The available source is a subscription preview rather than the full article, so it does not identify the individual datasets, provide metric-by-metric results, report computational costs or uncertainty estimates, or give enough implementation detail for independent reproduction. As a computational benchmark study, the results show an advantage within the evaluated datasets and experimental setup, but the preview does not establish how the method would perform on new applications, other data types or different classification systems.

Subject of Research: Feature selection for multi-label text classification using multi-order relevance and label interaction.

Article Title: MFLRA: leveraging multi-order feature–label relevance and ambiguity with label interaction for feature selection in multi-label text classification

Article References: Verma, G., & Sahu, T. P. (2026). MFLRA: leveraging multi-order feature–label relevance and ambiguity with label interaction for feature selection in multi-label text classification. Multimedia Tools and Applications, 85(11), Article 818. https://doi.org/10.1007/s11042-026-21971-7

Image Credits: AI Generated

DOI: 10.1007/s11042-026-21971-7

Keywords: MFLRA, multi-label text classification, feature selection, feature–label relevance, label interaction, label ambiguity, Elastic Net, entropy weighting, CODAS, MLKNN, particle swarm optimization, leveraging