基于图约束和预聚类的主动学习算法在威胁情景感知中的研究

发布时间：2018-03-26 09:50

  本文选题：图约束　切入点：预聚类　出处：《计算机应用研究》2017年05期

【摘要】：针对现有的威胁感知算法对样本标注代价较大,并且在训练分类器时只使用已标注的威胁样本,提出了一种基于图约束和预聚类的主动学习算法。该算法旨在通过降低标注威胁样本的代价,并且充分利用未标注的威胁样本对训练分类器的辅助作用,训练出更好的分类器以有效地感知威胁情景。该算法用已标注的威胁样本集合训练分类器,从未标注的威胁样本集中挑选出最有价值的威胁样本,并对其进行标注,再将标注后的威胁样本加入已标注的样本集中,同时删去原来未标注样本集中的此样本,最后用新的已标注的威胁样本集重新训练分类器,直到满足循环条件终止。仿真实验表明,基于图约束与预聚类的主动学习算法在达到目标准确率的同时降低了标注代价且误报率较低,能够有效地感知威胁情景,具有一定的研究意义。
[Abstract]:For the existing threat awareness algorithms, the cost of sample tagging is high, and only tagged threat samples are used in training classifier. This paper proposes an active learning algorithm based on graph constraint and preclustering, which aims to reduce the cost of tagging threat samples and make full use of unlabeled threat samples to assist the training classifier. A better classifier is trained to perceive the threat situation effectively. The algorithm uses the labeled threat sample set to train the classifier, and selects the most valuable threat sample from the untagged threat sample set and annotates it. Then the labeled threat sample is added to the labeled sample set, and the original unlabeled sample set is deleted. Finally, the new tagged threat sample set is used to retrain the classifier. The simulation results show that the active learning algorithm based on graph constraint and preclustering not only achieves the accuracy of target but also reduces the tagging cost and the false alarm rate is low. Has certain research significance.
【作者单位】： 南京南瑞集团公司/国网电力科学研究院;
【基金】：企业自选科技资助项目
【分类号】：TP181;TP309
，

本文编号：1667387