三种用于垃圾网页检测的随机欠采样集成分类器

发布时间：2018-06-02 04:54

本文选题：垃圾网页检测 + 不平衡分类　；参考：《计算机应用》2017年02期

【摘要】：针对垃圾网页检测过程中轻微的不平衡分类问题,提出三种随机欠采样集成分类器算法,分别为一次不放回随机欠采样(RUS-once)、多次不放回随机欠采样(RUS-multiple)和有放回随机欠采样(RUS-replacement)算法。首先使用其中一种随机欠采样技术将训练样本集转换成平衡样本集,然后对每个平衡样本集使用分类回归树(CART)分类器算法进行分类,最后采用简单投票法构建集成分类器对测试样本进行分类。实验表明,三种随机欠采样集成分类器均取得了良好的分类效果,其中RUS-multiple和RUS-replacement比RUS-once的分类效果更好。与CART及其Bagging和Adaboost集成分类器相比,在WEBSPAM UK-2006数据集上,RUS-multiple和RUS-replacement方法的AUC指标值提高了10%左右,在WEBSPAM UK-2007数据集上,提高了25%左右;与其他最优研究结果相比,RUS-multiple和RUS-replacement方法在AUC指标上能达到最优分类结果。
[Abstract]:In order to solve the problem of slight unbalance classification in the process of garbage page detection, three kinds of random under-sampling ensemble classifier algorithms are proposed, which are single unretractable random under-sampling (RUS-onceg), multiple unretractable random under-sampling (RUS-multiple) and random under-sampling (RUS-replacementation) algorithm. Firstly, one of the random under-sampling techniques is used to transform the training sample set into a balanced sample set, and then each balanced sample set is classified by using the classifying regression tree / cart classifier algorithm. Finally, a simple voting method is used to construct an integrated classifier to classify test samples. The experimental results show that all of the three random under-sampling ensemble classifiers have good classification effect, and RUS-multiple and RUS-replacement have better classification effect than RUS-once. Compared with CART and its Bagging and Adaboost integrated classifiers, the AUC index value of RUS-multiple and RUS-replacement methods on WEBSPAM UK-2006 data sets is increased by about 10%, and on WEBSPAM UK-2007 data sets by about 25%. Compared with other optimal results, the RUS-multiple and RUS-replacement methods can achieve the optimal classification results on the AUC index.
【作者单位】：南昌大学信息工程学院;南昌大学软件学院;
【基金】：江西省科技支撑计划项目(20131102040039)~~
【分类号】：TP393.092

【参考文献】