基于标签相关性的预测调整算法  

Prediction Adjustment Algorithm Based on Label Correlation

作　　者：张海涛 王丹东[1] 钱坤 闵帆[1] ZHANG Haitao;WANG Dandong;QIAN Kun;MIN Fan(School of Computer Science,Southwest Petroleum University,Chengdu 610500,China)

机构地区：[1]西南石油大学计算机科学学院,成都610500

出　　处：《海南热带海洋学院学报》2023年第5期72-81,共10页Journal of Hainan Tropical Ocean University

基　　金：国家社会科学基金后期资助项目(22FZXB092);中央引导地方科技发展项目(2021ZYD0003)。

摘　　要：多标签学习已成功应用于文本分类、图像识别等各个领域。流行的技术包括提取标签特定特征、利用标签相关性等。提出带有标签相关性的预测调整算法PALC(Prediction adjusting with label correlation)将标签相关性融入串行并行神经网络。一方面,采用新颖的、更有效的串行并行神经网络架构来替代常见的显式特征提取或压缩感知方法;另一方面,考虑用固有的标签矩阵内的相关性来计算相关性矩阵,并以流形正则的方式优化分类器。对10个基准数据集与7种流行算法进行比较,结果表明PALC在3大排名指标下均有优势。Multi-label learning has been successfully applied in various domains such as text classification and image recognition.Popular techniques include extracting label-specific features and exploiting label correlation.A Prediction Adjusting with Label Correlation(PALC)algorithm was proposed to incorporate label correlation into serial-parallel neural networks.On the one hand,a new and more effective serial-parallel neural network architecture was employed to replace the common explicit feature extraction or compressed sensing methods.On the other hand,the intrinsic correlation within the label matrix was considered to compute a correlation matrix,and the classifier was optimized by using manifold regularization.The currently conducted experiment involved a comparison of PALC with seven popular algorithms on ten benchmark datasets.The results,which were evaluated by using three ranking-based measures,consistently demonstrate the superiority of PALC across diverse domains.

关 键 词：标签相关性 流形正则化 多标签学习 串行并行神经网络 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]