基于半监督学习的数据流分类算法.docVIP

基于半监督学习的数据流分类方法 黄树成1,2 朱宇光2 董逸生1 1(东南大学计算机科学与工程学院，南京210096) 2(常州工学院计算机科学与工程系，江苏常州 213006) (schuang6@) 摘要 在数据流上建立有效的分类模型具有许多应用。流数据的动态性给业界提出了两个关键问题：如何监测数据流的变化；一旦发生显著变化,如何高效地组织足够的训练数据,调整无效的模型。本文提出了一个基于半监督学习的分类算法，较好地解决了这些问题。设计了一种识别显著变化的可靠方法；提出了一个基于Na?ve Bayes的EM算法，利用较少的类标数据来扩大训练数据集，从而极大地降低类标数据的需求量。基于新的训练数据可以构建一个有效的分类器。实验结果证明了算法的优势。 关键词 数据流分类；显著变化；类标数据；半监督学习；基于Na?ve Bayes的EM算法。 中图法分类号 TP311.13 Semi-supervised Learning Based Approach for Classifying Data Streams Huang Shucheng1,2 Dony Yisheng1 1(School of Computer Science and Engineering, Southeast University, Nanjing, 210096) 2(Department of Computer Science and Engineering, Changzhou Institute of Technology, Jiangsu Changzhou, 213006) Abstract Mining on data streams and constructing a classification model has found many applications. The dynamic nature of streaming data poses two key issues facing associated communities: one is how to monitor the potential changes occurring in data streams, and the other is how to collect sufficient labeled data to adjust the outdated model for adapting to significant changes. In this paper, a semi-supervised learning based algorithm is proposed to attack these problems. We devise a reliable method for monitor and differing significant changes from noisy changes. Whenever the significant changes present, an EM with na?ve Bayes algorithm is employed to augment currently insufficient labeled data into a sufficient training data set. So the need for labeled data is sharply reduced, and an effective classifier can be generated based on the new training data. Experiment results confirm the advantages of the algorithm. Keywords Data Streams Classification; Significant Changes; Labeled Data; Semi-supervised Learning; EM with Na?ve Bayes Algorithm.  1.引言 数据流挖掘是一个研究热点[1]。数据流的动态性给数据流分类提出了极大的挑战，关键是如何监测潜在的变化和组织新的训练数据响应数据流发生的变化。 许多数据流分类方法，比如[2，3，4]，假设类标数据容易获得且随时可以挖掘。但在实际应用中，数据的类标很难及时得到。 Wei F.等人在[5，6]中提出主动挖掘的思想，但检测变化的方法缺乏可靠性，而且模型的更新依赖于类标数据的数量。 我们提出了一种识别变化的方法和一种基于Na?ve Bayes的EM算法，可以降低更新分类模型