基于争议度的Boosting集成网络样本权值调整算法
来源期刊：中南大学学报(自然科学版)2012年第11期
论文作者：高敬阳 陈程立诏 朱群雄
文章页码：4355 - 4360
关键词：神经网络集成；差异度；精度；争议度
Key words：neural network ensemble; diversity; accuracy; standard deviation of Error-Right statistics
摘 要：神经网络集成AdaBoost算法权值调整策略对于分类正确或分类错误的样本采用统一的权值调整幅度，随着迭代次数的增加，统一的权值调整幅度将导致困难样本权重的过分积累，针对这一问题，提出基于争议度的权值调整策略，并采用的标准机器学习数据库UCI进行仿真实验。实验结果表明：该策略能够在样本权值修正阶段对各训练样本权值进行有区别的修改，即将多次连续分类错误的样本的权值提高幅度进行抑制，在一定程度上避免了困难样本权值过大而导致集成网络泛化性能下降，从而使得各个体分类器在不损失差异度的前提下获得理想的精度，提升集成网络的泛化性能，并具有良好的稳定性。
Abstract: Traditional AdaBoost algorithm sets difficult samples with too much weight and the over weighted difficult samples will lead to a declination of the ensemble performance, therefore, a new variant of AdaBoost called ERstd-AdaBoost algorithm were proposed. The experiments on several benchmark real-world sets available from the UCI repository were carried out. The results show that the strategy can update the weight of samples in the fixed stage, that is, it controls the increase of weight of repetitions misclassified samples. ERstd-AdaBoost algorithm can avoid setting misclassified samples too much weight which would lead to a declination of the ensemble performance. Hence, this new strategy can update difficult sample’s weight to achieve a better accuracy with no decline of diversity. The new algorithm can absolutely improve the performance of AdaBoost, and its stabilization is acceptable.
高敬阳，陈程立诏，朱群雄
(北京化工大学 信息科学与技术学院，北京，100029)
摘 要：神经网络集成AdaBoost算法权值调整策略对于分类正确或分类错误的样本采用统一的权值调整幅度，随着迭代次数的增加，统一的权值调整幅度将导致困难样本权重的过分积累，针对这一问题，提出基于争议度的权值调整策略，并采用的标准机器学习数据库UCI进行仿真实验。实验结果表明：该策略能够在样本权值修正阶段对各训练样本权值进行有区别的修改，即将多次连续分类错误的样本的权值提高幅度进行抑制，在一定程度上避免了困难样本权值过大而导致集成网络泛化性能下降，从而使得各个体分类器在不损失差异度的前提下获得理想的精度，提升集成网络的泛化性能，并具有良好的稳定性。
关键词：神经网络集成；差异度；精度；争议度
GAO Jing-yang, CHEN Cheng-li-zhao, ZHU Qun-xiong
(College of Information Science & Technology, Beijing University of Chemical Technology, Beijing 100029, China)
Abstract:Traditional AdaBoost algorithm sets difficult samples with too much weight and the over weighted difficult samples will lead to a declination of the ensemble performance, therefore, a new variant of AdaBoost called ERstd-AdaBoost algorithm were proposed. The experiments on several benchmark real-world sets available from the UCI repository were carried out. The results show that the strategy can update the weight of samples in the fixed stage, that is, it controls the increase of weight of repetitions misclassified samples. ERstd-AdaBoost algorithm can avoid setting misclassified samples too much weight which would lead to a declination of the ensemble performance. Hence, this new strategy can update difficult sample’s weight to achieve a better accuracy with no decline of diversity. The new algorithm can absolutely improve the performance of AdaBoost, and its stabilization is acceptable.
Key words:neural network ensemble; diversity; accuracy; standard deviation of Error-Right statistics
