融合样本相似性的弱监督多标签分类

Weakly Supervised Multilabel Classification Combining Sample Similarity

针对面向实际应用场景中数据标签易残缺导致有监督多标签分类方法可用训练数据量减少,未能利用大量标签缺失数据中蕴含的样本特征空间关联知识以最大化判别间隔,限制多标签分类效果等问题,本文提出一种融合样本相似性的弱监督多标签分类方法.该方法利用标签相关性和样本相似性恢复标签以提高数据利用率,并将标签恢复嵌入到训练过程中以便挖掘标签相关性,通过近端加速梯度法进行参数优化,建立弱监督学习场景的多标签分类模型.在真实数据集上的实验结果表明,该方法能够利用样本相似性有效提升模型在标签残缺时的分类能力,实用价值大.

Multilabel classification is a machine learning method to improve the performance of multi label joint decision by label correlation.In practical application scenarios,data labels are easy to be incomplete,which can lead to the reduction of available training data,and it is difficult to train the model adequately.Moreover,it is easy to cause the increase of label distribution variance,the deviation of correlation knowledge,and the limitation of multi label classification effect.To solve the problems,a weak supervised multi label classification method based on sample similarity was proposed.The method was arranged to use label correlation and sample similarity to recover labels to improve data utilization,and to embed label recovery into the training process to correct the bias in the model learning process.Based on the proximal accelerated gradient method,parameter optimization was carried out,and a multi label classification model was established for weak supervised learning scene.Experiments were completed with real data set.The results show that the method can effectively improve the classification ability of the model for the incomplete labels according to the similarity of samples,possessing high practical value.