Chen Suhua, Fang Xu, Zhai Feng, Wang Li, Lv Lin
School of Electrical Engineering, Xuchang University, Xuchang, 461000, Henan, China.
Henan Xj Metering Co., Ltd., Xuchang, 461000, Henan, China.
Sci Rep. 2025 May 13;15(1):16508. doi: 10.1038/s41598-025-01092-1.
Multi-label classification is a significant challenge in machine learning, especially as the dimensionality of the problem increases. As the number of dimensions grows, the performance of traditional classification algorithms often degrades substantially. Feature selection is a key technique for reducing dimensionality in multi-label scenarios, operating as a non-parametric process. Despite its importance, feature selection remains a complex issue without straightforward solutions, and various approaches using AI and evolutionary algorithms have been proposed to tackle it. However, these methods typically suffer from reduced efficiency and slower convergence as the dimensionality increases, due to the expanding search space. To address this issue and enhance convergence speed, this article introduces a hybrid AI solution that combines a binary particle swarm optimization algorithm with a local search strategy specifically designed for multi-label feature selection. Within this local search strategy, feature fusion plays a crucial role, where features are merged based on their relevance and correlation with the problem's output. These fused features are divided into two categories: those directly associated with the problem class and those that are similar to the problem class but distinct from other feature fusions. By leveraging this categorization, the particle swarm optimization technique is augmented with a local operator that removes redundant feature fusions and refines each solution. By incorporating this operator, the proposed method achieves superior convergence speed compared to previous algorithms in the field. The performance of the proposed method was evaluated across several datasets against some of the most widely used feature fusion selection algorithms. The experimental results demonstrated the proposed method's accuracy and efficiency, validating its effectiveness in multi-label feature selection.
多标签分类是机器学习中的一项重大挑战，尤其是随着问题维度的增加。随着维度数量的增长，传统分类算法的性能往往会大幅下降。特征选择是在多标签场景中降低维度的关键技术，它是一个非参数过程。尽管其很重要，但特征选择仍然是一个复杂的问题，没有直接的解决方案，并且已经提出了各种使用人工智能和进化算法的方法来解决它。然而，由于搜索空间不断扩大，这些方法通常会随着维度的增加而效率降低且收敛速度变慢。为了解决这个问题并提高收敛速度，本文介绍了一种混合人工智能解决方案，该方案将二进制粒子群优化算法与专门为多标签特征选择设计的局部搜索策略相结合。在这种局部搜索策略中，特征融合起着关键作用，其中特征是根据它们与问题输出的相关性和关联性进行合并的。这些融合特征分为两类：与问题类别直接相关的特征和与问题类别相似但与其他特征融合不同的特征。通过利用这种分类，粒子群优化技术通过一个局部算子得到增强，该算子可以去除冗余特征融合并优化每个解决方案。通过纳入这个算子，所提出的方法与该领域以前的算法相比实现了更高的收敛速度。在所提出的方法的性能在几个数据集上与一些最广泛使用的特征融合选择算法进行了评估。实验结果证明了所提出的方法的准确性和效率，验证了其在多标签特征选择中的有效性。
