基于最小二乘法的标记分布学习

摘要

多标记学习在一定程度上解决了标记多义性问题,它主要关注实例对应的相关标记或者无关标记,而标记分布能够反映相关标记对于实例的重要程度.从重构标记分布的思想出发,利用最小二乘法建立模型,提出基于最小二乘法的标记分布学习(lsm-LDL).首先用特征重构标记,通过变换矩阵使得每一个标记能够表示为特征的一个线性组合;然后用最小二乘法建立优化模型;最后引入L2范数规则化项,防止过拟合,保证泛化能力.在4个实际的数据集上进行实验,并与3种已有的标记分布学习算法在5种评价指标上进行比较,实验结果表明提出的lsm-LDL算法是有效的.%The importance of the labels relative to the instance can be reflected by label distribution . Multi-label learning could solve ambiguity problems of label by focusing on the corresponding related or unrelated labels of the instance .The label distribution learning based on least square method ( lsm-LDL) was proposed .Firstly, Some features were used to reconstruct the label , and then the transformation ma-trix was used to have each label expressed as a linear combination of features .Secondly , the least square method was applied to establish the optimization model .Finally, the L2 norm regularization term was in-troduced to prevent overfitting , and to ensure the generalization ability .Experiments were carried out on four actual data sets , and the lsm-LDL algorithm was compared with three other existing labeled distribu-tion learning algorithms with five evaluation indices .The results showed that the proposed lsm-LDL algo-rithm was effective .