《Table 2 Results of five experiments under different models》

  提示：宽带有限、当前游客访问压缩模式

本系列图表出处文件名：随高清版一同展现

《基于XGBoost算法的异常用户识别（英文）》

1. 获取 高清版本忘记账户？点击这里登录

1. 下载图表忘记账户？点击这里登录

The balanced sample set was composed of 1 394positive samples and 1 394negative samples.The former samples were the existing abnormal users，and the latter samples were the samples randomly selected from 8 562 normal users.The balanced sample set was divided into five parts，four of which were training sets，one of which was a test set.The experiments were conducted five times under different recognition models.KNN classifier taked K=50.In the BP neural network，the number of hidden layer unit was 8.The training algorithm used QuickProp.The algorithm parameters were 0.1，2，0.000 1and 0.1.The maximum number of iterations was 1 000.The XGBoost parameters were divided into three kinds:general parameters，booster parameters and learning target parameters.In this experiment，0.01 was selected as the booster parameter shrinkage step size（ETA）to prevent overfitting，and the maximum iteration number（nrounds）was 1 500.The minimum sample weight of child nodes（min_child_weight）was 10.The ratio of feature sampling（colsample_bytree）was 0.8.In the learning target parameters，the objective function selected binary logistic regression（binary:logistic），and the evaluating indicator was the average accuracy（map）.The rest parameters kept the default values.The results are shown in Table 2and the values are the average of the five experiments.