摘要
在多标签分类问题中,标签之间往往是相关的,为了提高分类性能,利用标签之间的相关性,提出AdaBoost.MLR算法和标签相关性分析方法。AdaBoost.MLR算法采用余弦相似度来计算标签相关性矩阵,利用标签相关性矩阵对原始标签矩阵进行补全转换为模糊标签矩阵,将标签空间划分为标签集、标签相关集和标签无关集,结合标签之间的相关性和弱分类器的分类情况,对样本权重进行调整。AdaBoost.MLR算法也能解决多类别分类问题,在其标签相关性的计算中,根据已经训练的弱分类器得到的临时强分类器的分类结果,构造标签相似性矩阵。实验结果表明,文中提出的算法在实验数据集上优于现有的算法,尤其在标签相关性复杂的数据集上分类性能有显著提升。
In order to improve classification performance and exploit label correlations, AdaBoost. MLR algorithm was proposed. Cosine similarity was adopted to capture the complex correlations among labels in AdaBoost. MLR algorithm, a supplementary label matrix was incorporated, which augments the incomplete original label matrix by exploiting the label correlations, label space was divided into three parts of label set, relevant label set and irrelevant label set, weight-update rule was modified according to correlations among labels and the results of weak learner. AdaBoost. MLR algorithm was able to solve multi-class classification problem specially,label similarity ma- trix,instead of cosine similarity, was constructed by the classification results of temporary strong learner combined by previous trained weak learners. The experimental results illustrated that the proposed algorithm was superior to existing algorithms, and the classification performance was improved significantly on datasets had complex correlations among labels.
出处
《四川大学学报(工程科学版)》
EI
CAS
CSCD
北大核心
2016年第5期91-97,共7页
Journal of Sichuan University (Engineering Science Edition)
基金
四川省科技支撑计划资助项目(2012GZ0106)
中科院西部之光人才培养计划资助项目
四川省科技创新苗子工程资助项目(2015060)
关键词
标签相关性
多标签分类
多分类问题
ADABOOST算法
分类器组合
ensemble learning
label correlation
multi-label classification
multi-class classification
AdaBoost algorithm
classifier combination