一种面向多标签分类的在线主动学习算法

An online active learning algorithm for multi-label classification

针对现有算法多标签分类器收敛效率低和标签查询策略未考虑特征辨别能力的弊端,提出一种基于判别采样和镜像梯度下降规则的多标签在线主动学习算法(multi-label active mirror descent by discrimination sampling,MLAMD_D)。MLAMD_D算法采用二元关联策略将包含C个标签的多标签分类问题分解成C个相互独立的二分类问题,算法使用镜像梯度下降规则更新其二分类器,并采用基于判别的采样策略。将MLAMD_D算法与现有算法以及基于随机采样和镜像梯度下降规则的多标签在线主动学习算法(multi-label active mirror descent by random sampling,MLAMD_R)在6个多标签分类数据集上进行对比试验。试验结果表明,MLAMD_D算法的多标签分类性能优于其他多标签在线主动学习算法。因此,MLAMD_D算法在处理多标签在线主动学习的任务中具有可行性和有效性。

Multi-label active mirror descent by discrimination sampling(MLAMD_D)was proposed to overcome the drawbacks of the existing algorithms,such as low convergence efficiency of multi-label classifier and the label query strategy did not consider the discriminative ability of features.The MLAMD_D algorithm used the binary relevance strategy to decompose the multi-label classification problem with C labels into C independent binary classification problems.The MLAMD_D algorithm used the mirror descent rule to update each binary classifier,meanwhile,the algorithm adopted the discrimination-based sampling strategy.The proposed algorithm was compared with the existing algorithms and multi-label active mirror descent by random sampling algorithm(MLAMD_R)on 6 multi-label classification data sets.The convincing experimental results showed that the performance of the MLAMD_D algorithm was superior to other multi-label online active learning algorithms.Therefore,the MLAMD_D algorithm was feasible and effective when dealing with multi-label online active learning tasks.