A framework for classifier-neutral ensemble multi-label feature selection via swarm optimization

• An ensemble feature selection method is proposed for multi-label classification. • Rank aggregation is formulated as an optimization problem based on Spearman’s footrule distance. • The problem is solved using particle swarm optimization. • The proposed method reduces computational cost by avoiding learning algorithm dependence. • Experimental results on 10 datasets confirm superior performance over base methods. Multi-label classification, wherein many labels are assigned to a single instance simultaneously, has become increasingly prevalent in real-world applications. With the rapid expansion of data in recent years, the need for effective feature selection algorithms has become more pronounced. Notably, ensemble feature selection algorithms provide the possibility of balancing the strengths and drawbacks of multiple feature selection algorithms. Concurrently, swarm intelligence-based feature selection methods gained attention for their ability to efficiently navigate large and complex search spaces. However, a significant limitation of many existing methods is their reliance on classifier-dependent evaluation criteria. In this study, we present a new framework for ensemble-based multi-label feature selection that is entirely independent of any specific classifier. The particle swarm optimization (PSO) is taken as the representative algorithm, and the rank aggregation problem is formulated as an optimization task. The objective function is designed to minimize the total Spearman’s footrule distance between the aggregated and the individual feature selection rankings. We evaluate the method on ten multi-label datasets with six widely accepted evaluation metrics. The experimental findings reveal the superior performance and robustness. Additionally, the framework can be readily extended to other swarm intelligence algorithms.