摘要
针对标签特定特征多标签学习算法(multi-label learning with label-specific features,LIFT)未能在聚类以及分类阶段考虑标签相关性问题,提出一种基于标签相关性的标签特定特征多标签学习算法(multi-label learning with label-specific features via label correlations,LFLC).将标签空间加入特征空间进行聚类构建分类模型,采用考虑标签相关性的聚类集成技术为每个标签构造标签特定特征,使用相关性矩阵构建无向完全图并挖掘图中标签集合相关性,通过树集成表达标签间多种不同结构的强相关性.在试验部分,采用涵盖不同领域的10个数据集,以Hamming Loss、Ranking Loss、One-error、Coverage、Average Precision和macroAUC为评估指标,进行了参数敏感性分析和统计假设检验.结果表明:结合聚类集成与标签间强相关性的LFLC算法较其他对比多标签算法整体上能取得较好的效果.
To solve the problem that multi-label learning with label specific features(LIFT)could not consider label correlation in the clustering and classification stages,a method for multi-label learning with label-specific features via label correlations(LFLC)was proposed.The label space was added to the feature space for clustering to construct the classification model,and the clustering ensemble with considering label correlation was used to construct label-specific features for each label.The correlation matrix was used to construct undirected complete graph and mine the correlation of label sets in the graph.The strong correlation of multiple different structures between labels was expressed by tree ensemble.In the experiment,10 data sets covering different fields were used,and Hamming Loss,Ranking Loss,One-error,Coverage,Average Precision and macroAUC were used as evaluation indexes to carry out parameter sensitivity analysis and statistical hypothesis test.The results show that the LFLC algorithm combined with clustering ensemble and strong correlation between labels can obtain better performance generally.
作者
王进
梁晨
孙开伟
陈乔松
邓欣
WANG Jin;LIANG Chen;SUN Kaiwei;CHEN Qiaosong;DENG Xin(Key Laboratory of Data Engineering and Visual Computing,Chongqing University of Posts and Telecommunications,Chongqing 400065,China)
出处
《江苏大学学报（自然科学版）》
CAS
北大核心
2023年第5期554-563,576,共11页
Journal of Jiangsu University：Natural Science Edition
基金
国家自然科学基金资助项目(61806033)。
关键词
多标签学习
标签特定特征
聚类集成
标签相关性
无向完全图
最小生成树
multi-label learning
label-specific feature
clustering ensemble
label correlation
undirected complete graph
minimum spanning tree
