在插入符号（R）中分配权重参数时出现错误“可变长度不同”

Question

我想使用下面的代码在插入符号中应用加权观察：

model_weights <- ifelse(train$y == 0,
                        (1/table(train$y)[1]) * 0.5,
                        (1/table(train$y)[2]) * 0.5)

xgbT <- train(x = as.matrix(train[,-21]), y = make.names(as.factor(train$y)), 
              method = "xgbTree", 
              trControl = cctrl1,
              metric = "MCC",
              maximize = TRUE,
              weights = model_weights,
              preProc = c("center", "scale"),
              tuneGrid = expand.grid(nrounds = c(150), #number of trees
                                    max_depth = c(7), #max tree depth
                                    eta = c(0.03), #learning rate
                                    gamma = c(0.3), #min split loss
                                    colsample_bytree = c(0.7),
                                    min_child_weight = c(10, 1, 5), #min number of instances in the leaf
                                    subsample = c(0.6)), #subsample ratio of the training instance
              early_stop_round = c(3), #if no improvements over specified rounds
              objective = c("binary:logistic"),
              silent = 0)

但是，它给了我这个错误：Error in model.frame.default(formula = .outcome ~ ., data = dat, weights = wts) : variable lengths differ (found for '(weights)')

虽然我已经检查过它们的长度与下面的代码相同：

> table(model_weights)
model_weights
0.0000277654375832963  0.000231481481481481 
                18008                  2160 
> table(train$y)

    0     1 
18008  2160 

知道如何解决这个问题吗？

注意：我可以在没有 weights 参数的情况下运行 train 函数。

Answer 1

经过进一步调试，发现问题是因为我在cctrl1中应用了sampling。因此，weights 的长度不同，因为我在它们应用重新采样之前生成它。

因此，您只需从 trControl 中删除 sampling 即可解决此问题。如果您仍然想应用重新采样，那么您必须在运行以下代码之前重新采样数据：

model_weights <- ifelse(train$y == 0,
                    (1/table(train$y)[1]) * 0.5,
                    (1/table(train$y)[2]) * 0.5)