基于正交约束和最大类内特征判别性的分层分类特征选择算法  

Hierarchical Classification Feature Selection Algorithm Based on Orthogonal Constraints and Intra-class Maximum Feature Discriminability

作　　者：殷金钊 郑文利 钱婷[2,3] 折延宏[2,3] YIN Jinzhao;ZHENG Wenli;QIAN Ting;SHE Yanhong(College of Computer,Xi'an Shiyou University,Xi'an 710065,China;College of Science,Xi'an Shiyou University,Xi'an 710065,China;Institute of Concepts,Cognition and Intelligence,Northwest University,Xi'an 710127,China)

机构地区：[1]西安石油大学计算机学院,陕西西安710065 [2]西安石油大学理学院,陕西西安710065 [3]西北大学概念、认知与智能研究中心,陕西西安710127

出　　处：《山西大学学报(自然科学版)》2025年第1期144-157,共14页Journal of Shanxi University(Natural Science Edition)

基　　金：国家自然科学基金(12171388,61976244,12101478,12171294);陕西省自然科学基础研究计划项目(2023JCYB027);浙江海洋大学海洋大数据挖掘与应用重点实验室(OBDMA202101);陕西数理基础科学研究项目(23JSZ008);研究生创新项目(YCX2413143)。

摘　　要：在大数据时代,数据正呈现出指数级增长趋势。数据间的类别层次结构使得分类学习任务更有效率。现有的分层分类特征选择算法未充分体现出类内特征的判别性,因此本文提出了一种基于正交约束和最大化类内特征判别性的分层分类特征选择算法(Hierarchical Classification Feature Selection Algorithm Based on Orthogonal Constraints and Intra-class Maximum Feature Discriminability,HFSOC)。该算法在使用稀疏正则化项去除不相关特征后,利用改进后的正交约束公式来度量类间独立性,并将每个内部节点特征矩阵的各个列向量互相正交,以提高类内特征的判别性。最后,利用递归正则项优化输出特征权重矩阵。实验结果表明,本文所提算法在5个数据集上取得了一定的效果,其分类准确率在DD数据集上相比于HFisher算法提高约17%,在F194数据集和CLEF数据集上相比于基于l_(2,1)范数最小化的高效鲁棒的特征选择算法(HFSNM)均提高约10%,在ILSVRC数据集上相比于HFSNM算法提高约1%。In the era of big data,data is showing an exponential growth trend.The hierarchical structure of categories among data makes the classification learning task more efficient.However,existing hierarchical classification feature selection algorithms do not fully reflect the discriminative nature of intra-class features.This paper proposes a hierarchical classification feature selection algorithm based on orthogonal constraints and maximizing the discriminative nature of intra-class features(HFSOC).The algorithm utilizes an improved orthogonality constraint formula to measure inter-class independence and orthogonalizes the individual column vectors of each internal node's feature matrix to each other to improve the discriminative property of intra-class features after using sparse regularization terms to remove irrelevant features.Finally,the output feature weight matrix is optimized using recursive regularization terms.The experimental results show that the algorithm proposed in this paper achieves certain results on five datasets,and its classification accuracy is improved by about 17%compared to HFisher's algorithm on the DD dataset,by about 10%compared to efficient and robust feature selection via joint l_(2,1)-norms minimization algorithm(HFSNM)on both the F194 dataset and the CLEF dataset,and by about 1%compared to HFSNM's algorithm on the ILSVRC dataset.

关 键 词：特征选择 稀疏学习 层次树结构 正交约束 递归正则化 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]