基于双层结构的多标签优序选择分类算法

Algorithm for multi-label classification with optimal sequence based on double layers

针对已有的多标签分类算法在设计过程中忽略标签之间关联性,导致分类精度降低的问题,提出基于双层结构的多标签优序选择分类(DLMC-OS)算法。通过二次信息交互实现标签间的关联性,解决链式分类模型随机性影响分类精度的问题。DLMC-OS构建一个双层结构的分类模型:第一层采用典型的二元独立分类模型实现对实例的第一次分类,与第二层进行标签信息的交互;第二层构建带有更新过程的链式分类模型,用链来传递和更新标签信息,实现分类信息的二次交互。提出构建具有最大权重的标签生成树(MWT-OS)算法,寻求标签优序,解决链式分类模型随机选择类标号序列训练二值分类器导致分类精度降低的问题。在9组基准数据集上与相关算法的比较验证了该算法的有效性。

Current multi-label classification algorithm ignores the correlation among labels during the process of design,leading to the low classification accuracy.A multi-label classification algorithm based on double-layer was proposed（DLMC-OS）.The label correlation was realized through secondary information exchange.The problem of classification accuracy caused by randomness of classification chain was also solved.A classification model DLMC-OS with double layers was built.The first layer adopted typical binary classification model to achieve independence in the first instance classification and interacted label information with the second layer.The second layer classification model built a model of classification chain with the ability of updating.The chain was utilized to transfer and renew label information to achieve secondary interaction for classified information.Also a building having a maximum weight of labels spanning（MWT-OS）was also proposed,seeking to label pecking order,solving the problem of reduced classification accuracy caused by randomly selecting class label sequence to train binary classifier by the chain classification model.Comparisons on nine benchmark data sets with related algorithms verify the effectiveness of the algorithm.