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

摘要

In view of the existing threat perception algorithm,the cost of sample labeling was large and just used the labeled samples to train the classifier,this paper put forward an activing learning algorithm,which was based on the the algorithm of graph constraints and pre-clustering to achieve a better classifier,by reducing the costs of threat sample labeling and making use of the auxiliary effects to train the classifier with unlabeled threat samples.Firstly,the algorithm used the labeled threat samples set to train the classifier,then picked the most valuable sample from unlabeled threat samples set and label ited,after that,put the labeled sample into the labeled sample set and deleted this sample from the original unlabeled sample set.Finally,it used this new set to retrain the classifier until meeting the conditions which was preset.Experiments show that it is effective to aware the threat situation which is based on this method,and it is reducing the cost of labeling and has a lower false alarm rate without loss accuracy at the same time.%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.