基于在线迁移学习的重现概念漂移数据流分类

发布时间：2018-01-13 11:39

本文关键词：基于在线迁移学习的重现概念漂移数据流分类　出处：《计算机研究与发展》2016年08期 　论文类型：期刊论文

【摘要】：随着大数据时代的到来,数据流分类被应用于诸多领域,如:垃圾邮件过滤、市场预测及天气预报等.重现概念是这些应用领域的重要特点之一.针对重现概念的学习与分类问题中的"负迁移"和概念漂移检测的滞后性,提出了一种基于在线迁移学习的重现概念漂移数据流分类算法——RC-OTL.RC-OTL在检测到概念漂移时存储刚学习的一个基分类器,然后计算最近的样本与存储的各历史分类器之间的领域相似度,以选择最适合对后续样本进行学习的源分类器,从而改善从源领域到目标领域的知识迁移.另外,RC-OTL还在概念漂移检测之前根据分类准确率选择合适的分类器对后续样本分类.初步的理论分析解释了RC-OTL为什么能有效克服"负迁移",实验结果进一步表明:RC-OTL的确能有效提高分类准确率,并且在遭遇概念漂移后能更快地适应后续样本.
[Abstract]:With the arrival of big data era, data stream classification has been applied in many fields, such as spam filtering. The concept of recurrence is one of the important characteristics of these application fields. The "negative migration" and the lag of the detection of concept drift in the learning and classification of the concept of recurrence are discussed in this paper. In this paper, an online migration learning based classification algorithm for reproducing conceptual drift data streams is proposed. RC-OTL.RC-OTL stores a base classifier for rigid learning when it detects concept drift. Then the domain similarity between the most recent samples and the stored historical classifiers is calculated to select the source classifier that is the most suitable for the subsequent samples to learn. Thus improving knowledge migration from source domain to target domain. RC-OTL also selects the appropriate classifier to classify the subsequent samples according to the classification accuracy rate before the concept drift detection. The preliminary theoretical analysis explains why RC-OTL can effectively overcome the "negative migration". The experimental results further show that: RC-OTL can effectively improve the classification accuracy and adapt to subsequent samples more quickly after the concept drift.
【作者单位】：桂林电子科技大学计算机与信息安全学院;广西可信软件重点实验室(桂林电子科技大学);广西信息科学实验中心(桂林电子科技大学);河北科技大学信息学院;
【基金】：国家自然科学基金项目(61363029,U1501252) 广西区自然科学基金项目(2014GXNSFAA118395) 广西区科学研究与技术开发项目(桂科攻14124005-2-1) 广西信息科学中心项目(YB408)~~
【分类号】：TP311.13
【正文快照】：

【相似文献】