Question

我是使用xgboost软件包的新手，我正在尝试创建精度最高的模型，并且使灵敏度和特异性达到平衡。我遇到的问题是，基数大约是1：3不平衡，而且预测使我的灵敏度非常低。

data(cats,package = "MASS")
prop.table(table(cats$Sex))
        F         M 
0.3263889 0.6736111

绘制数据

library(ggplot2)
ggplot(cats, aes(Bwt, Hwt, colour = Sex, shape = Sex)) + geom_point(size = 3)

加载xgboost软件包

library(xgboost)

# Split data
set.seed(123);index <- sample(nrow(cats),size = nrow(cats)*0.75)
train <- cats[index,]; test <- cats[-index,]

train_x <- train; test_x <- test
train_x$Sex <- NULL; test_x$Sex <- NULL

# Convert predict variable in numeric
y_train <- as.numeric(train$Sex) - 1; y_test <- as.numeric(test$Sex) - 1
train_x[] <- sapply(train_x, as.numeric); test_x[] <- sapply(test_x, as.numeric)

# Construct xgb.DMatrix object from either a dense matrix
dtrain <- xgb.DMatrix(as.matrix(train_x),label = y_train)
dtest <- xgb.DMatrix(as.matrix(test_x))

构建模型

## xgboost parameters
xgb_params <- list(booster = "gbtree" 
                , objectve = "binary:logistic"
                , eta=0.1 #default 0.3 [0,1]
                , gamma=0
                , max_depth=7 # default 6 Typical values: 3-10
                , subsample=1
                , tree_method = "exact"
                , scale_pos_weight = 5
                , base_score=median(y_train)
                , seed = 2018)

# tuning Cross Validation 
xgbcv <- xgb.cv(params = xgb_params
                , data = dtrain
                , nrounds = 2000
                , nfold = 7
                , print_every_n = 5
                , early_stopping_rounds = 40
                , maximize = F
                , prediction = F
                , showsd = T
                , metrics = "error")

# train model
gb_dt <- xgb.train(params = xgb_params
                , data = dtrain
                , nrounds = xgbcv$best_iteration
                , print_every_n = 2
                , early_stopping_rounds = 40
                , maximize = F
                , watchlist = list(train=dtrain))

test_probs <- predict(gb_dt, dtest, type = "response")
test_preds <- as.numeric(test_probs > .5)

# Change predicted values to match original data set, check accuracy
test_submit <- 0
test_submit[test_preds==0] <- "F"
test_submit[test_preds==1] <- "M"

我计算出混淆矩阵

caret::confusionMatrix(as.factor(test_submit), test$Sex)
Confusion Matrix and Statistics

          Reference
Prediction  F  M
         F  7  0
         M  7 22

               Accuracy : 0.8056          
                 95% CI : (0.6398, 0.9181)
    No Information Rate : 0.6111          
    P-Value [Acc > NIR] : 0.01065         

                  Kappa : 0.55            
 Mcnemar's Test P-Value : 0.02334         

            Sensitivity : 0.5000          
            Specificity : 1.0000          
         Pos Pred Value : 1.0000          
         Neg Pred Value : 0.7586          
             Prevalence : 0.3889          
         Detection Rate : 0.1944          
   Detection Prevalence : 0.1944          
      Balanced Accuracy : 0.7500          

       'Positive' Class : F

如何在不降低准确性的情况下提高模型的灵敏度？也许密钥在scale_pos_weight中，我不知道。任何建议都欢迎。 注意：我的兴趣是拥有一个最佳平衡的模型。尽可能正确地对最大数量的母猫进行分类。

Answer 1

首先使用max_depth = 6树的最大深度。现在使用以下代码：

library(caret)
k <- seq(0.5,0.99,0.01)
Sensitivity <- rep(0,length(k))

for(i in seq_along(k)){
  test_preds <- as.numeric(test_probs > k[i])
  test_submit <- 0
  test_submit[test_preds==0] <- "F"
  test_submit[test_preds==1] <- "M"
  tab <- confusionMatrix(as.factor(test_submit), test$Sex)
  Sensitivity[i] = tab$byClass[[1]]}

max(Sensitivity)
[1] 0.8571429
pos <- which.max(Sensitivity)

test_preds <- as.numeric(test_probs > k[pos])
test_submit <- 0
test_submit[test_preds==0] <- "F"
test_submit[test_preds==1] <- "M"

计算混淆矩阵

confusionMatrix(as.factor(test_submit), test$Sex)
Confusion Matrix and Statistics

          Reference
Prediction  F  M
         F 12  5
         M  2 17

               Accuracy : 0.8056          
                 95% CI : (0.6398, 0.9181)
    No Information Rate : 0.6111          
    P-Value [Acc > NIR] : 0.01065         

                  Kappa : 0.6063          
 Mcnemar's Test P-Value : 0.44969         

            Sensitivity : 0.8571          
            Specificity : 0.7727          
         Pos Pred Value : 0.7059          
         Neg Pred Value : 0.8947          
             Prevalence : 0.3889          
         Detection Rate : 0.3333          
   Detection Prevalence : 0.4722          
      Balanced Accuracy : 0.8149          

       'Positive' Class : F

正如我们现在看到的那样，您将拥有一个具有更高灵敏度并保持准确性的模型，因此您可以以更少的误差对雌猫进行分类。 注意：但是，价格会降低特异性。

训练xgboost模型不平衡数据

1 个答案: