基于动态异构集成的多标签数据流分类算法  

Multi-label data stream classification algorithm based on dynamic heterogeneous ensemble

作　　者：丁剑[1] 武红鑫 韩萌[1] DING Jian;WU Hong-xin;HAN Meng(School of Computer Science and Engineering,North Minzu University,Yinchuan 750021,China)

机构地区：[1]北方民族大学计算机科学与工程学院,宁夏银川750021

出　　处：《计算机工程与设计》2023年第10期3031-3038,共8页Computer Engineering and Design

基　　金：宁夏自然科学基金项目(2022AAC03279);国家自然科学基金项目(62062004);中央高校基本科研业务费专项基金项目(2022PT_S04)。

摘　　要：使用反映数据变化机制的自适应模型可以更好处理数据流问题。为实现自适应调整集成分类器使其更符合数据特性,提出一种基于动态异构集成的多标签数据流分类算法。通过使用H个不同分类算法分别训练固定大小的数据块,生成候选分类器组E={E 1,…,E H},利用几何加权公式计算每个E i中候选基分类器的权重实现组内的动态更新;提出一种新的自适应选择策略生成最终的异构集成分类器。通过在6个数据集上的大量实验验证,提出算法比现有算法在准确度、基于实例的F1值、微观F1值、宏观F1值上有更好的性能。Data streams can be better handled using adaptive models that reflect data change mechanisms.To implement an adaptive adjustment of the ensemble classifiers to make it more consistent with the data characteristics,a multi-label data stream classification algorithm based on dynamic heterogeneous ensemble(DHEML)was proposed.Candidate classifier sets E={E 1,…,E H}were generated by training fixed-size data blocks separately using H different classification algorithms,and a geometric weighting formula was used to calculate the weights of candidate base classifiers in each E i to achieve dynamic updates within the group.A new adaptive selection strategy was proposed to generate the final heterogeneous ensemble classifier.Through extensive experiments on six datasets,it is demonstrated that DHEML has better performance than existing algorithms in terms of accuracy,instance-based F1 values,micro-F1 values,and macro-F1 values.

关 键 词：多标签分类 数据流 异构 集成分类器 动态更新 自适应选择策略 几何加权 

分 类 号：TP393[自动化与计算机技术—计算机应用技术]