摘要： 针对多标记学习分类问题，算法适应方法将其转化为排序问题，并将输出标记按照其与示例的相关性进行排序，该类方法取得了较好的分类效果。基于间隔准则提出一种多标记学习算法，通过优化模型在示例的相关标记集合中最小输出与不相关标记集合中最大输出的间隔损失来进行标记排序。在此基础上，为充分利用全部标记信息，提出一种改进的优化排序多标记学习算法，分别优化模型在示例的相关标记集合中平均输出与不相关标记集合中最大输出的间隔损失，以及优化模型在相关标记集合中最小输出与不相关标记集合中平均输出的间隔损失，从而实现标记排序。在模型的参数学习过程中，使用改进的次梯度Pegasos算法进行优化。将所提2种算法与ML-RBF、BP-MLL、ML-KNN多标记学习算法在4个多标记数据集上进行对比实验，结果表明，在HL、RL等5种不同的评价准则下，2种算法均能与对比算法取得相近的分类性能。

关键词: 多标记学习, 算法适应, 标记排序, 平均输出, 间隔准则, Pegasos算法

Abstract: For classification problems in multi-label learning,the algorithm adaptation methods that transform them into a ranking problem and rank the output labels according to their relevance to the examples have made great success.This paper proposes a multi-label learning algorithm based on the margin criterion,which optimizes the margin loss between the minimum output in the relevant label set of examples and the maximum output in the irrelevant label set of examples,so as to sort the labels.On this basis,in order to utilize all the label information,an improved optimized ranking algorithm for multi-label learning is proposed to respectively optimize the margin loss between the average output in the relevant label set and the maximum output in the irrelevant label set of examples,and the margin loss between the minimum output in the relevant label set and the average output in the irrelevant label set,so as to sort the labels.Then an improved sub-gradient Pegasos algorithm is used to learn the model parameters.Experimental results on four multi-label datasets show that the two improved algorithms achieves similar classification performance compared with ML-RBF,BP-MLL,and ML-KNN under HL,RL and other three different evaluation criteria.

Key words: multi-label learning, algorithm adaptation, label ranking, average output, margin criterion, Pegasos algorithm

中图分类号: 

• TP391