使用tidymodels调优Ames房价模型时遇下标越界错误求助

问题背景 #

使用tidymodels处理Kaggle Ames房价数据集时，运行模型调优代码出现下标越界错误，具体信息：

x Fold1: preprocessor 1/1, model 1/12: Error in y.mat[, 2]: subscript out of bounds
Warning: All models failed. Run show_notes(.Last.tune.result) for more information.

执行show_notes后仍仅返回相同错误提示。

我的代码流程 #

1. 数据拆分与预处理Recipe #

data.split = initial_split(house.prices, prop=0.8)

train.data = training(data.split)
test.data = testing(data.split)

# Create recipe - OLS
sales.rec = recipe(SalePrice ~., data = train.data) %>%
  step_log(SalePrice, LotArea, GrLivArea, TotRmsAbvGrd) %>%
  update_role(Id, SalePrice, new_role = "ID") %>%
  step_num2factor(MSSubClass, levels = as.character(unique(house.prices$SalePrice))) %>%
  step_unknown(PoolQC, Fence, MiscFeature, BsmtQual, BsmtCond, BsmtExposure, BsmtFinType1,
               FireplaceQu, GarageType, GarageQual, GarageCond, new_level = "None") %>%
  step_mutate(PorchArea = OpenPorchSF+EnclosedPorch+`3SsnPorch`+ScreenPorch) %>%
  step_mutate(garage.age = YrSold - GarageYrBlt,
         house.age = YrSold - YearBuilt,
         renovation.age = YrSold - YearRemodAdd
         ) %>%
  step_mutate(has.garage = (GarageType != NA),
         has.basement = (BsmtExposure != NA),
         has.pool = (PoolQC != NA),
         is.new = house.age==0
         ) %>%
  step_cut(OverallQual, breaks = c(2.5, 6.5, 8.5)) %>%
  step_cut(OverallCond, breaks = c(2.5, 6.5, 8.5)) %>%
  step_mutate(house.age = log(1+house.age)) %>%
  step_rm(-c(SalePrice, house.age, renovation.age, 
             has.garage, has.basement, has.pool, is.new,
             MSZoning, LotArea, Alley, LotShape, Utilities, Neighborhood, OverallQual, OverallCond, 
             ExterCond, Foundation,
             BsmtFinSF1, Heating, HeatingQC, CentralAir, GrLivArea, FullBath, KitchenQual, TotRmsAbvGrd,
             PavedDrive, PorchArea, MiscVal,SaleCondition)) %>% 
  step_other(all_nominal_predictors(), all_factor(), all_string()) %>%
  step_string2factor(all_string_predictors()) %>%
  step_nzv(all_predictors()) %>%
  step_impute_median(all_numeric_predictors()) %>%
  step_unknown(all_factor_predictors()) %>%
  step_normalize(all_numeric_predictors())

2. 随机森林模型搭建 #

my.rf = rand_forest(mtry = tune(), trees = 2000, min_n=tune()) %>%
  set_engine("ranger") %>%
  set_mode("regression")

3. 工作流与模型调优 #

tree.grid = expand.grid(min_n=c(2,14,27,40), mtry = c(4, 8, 12))
folds = rsample::vfold_cv(train.data, v = 5)
metric = metric_set(rmse)


set_dependency("rand_forest", "ranger", "ranger", mode = "regression")
# Random forest
my.rf.rec = sales.rec %>%
  step_dummy(all_nominal_predictors()) %>%
  step_zv(all_predictors()) %>%
  step_other(all_nominal_predictors())
my.rf.wflow = workflow() %>%
  add_model(my.rf) %>%
  add_recipe(my.rf.rec)

my.rf.res = my.rf.wflow %>%
  tune_grid(
  resamples = folds,
  metrics = metric,
  grid = tree.grid
)

错误根源与修复方案 #

1. 核心错误：误将目标变量设为ID角色 #

在update_role(Id, SalePrice, new_role = "ID")中，把回归目标SalePrice标记为ID列，导致模型训练时无法找到目标变量，直接引发下标越界。

修复：仅将Id设为ID角色：

update_role(Id, new_role = "ID") %>%

2. step_num2factor的levels参数完全错误 #

用房价的唯一值作为房屋类型编码MSSubClass的因子水平，逻辑完全混乱。

修复：使用MSSubClass自身的唯一值：

step_num2factor(MSSubClass, levels = as.character(unique(train.data$MSSubClass))) %>%

3. NA判断方式错误 #

GarageType != NA这种写法在R中不成立（NA != NA返回NA），必须用!is.na()判断缺失值。

修复：

step_mutate(has.garage = !is.na(GarageType),
         has.basement = !is.na(BsmtExposure),
         has.pool = !is.na(PoolQC),
         is.new = house.age==0
         ) %>%

4. 重复调用step_other #

原始Recipe中已经处理过名义变量的低频类别，后续工作流中重复调用会导致冲突。

修复：移除my.rf.rec中的step_other(all_nominal_predictors())。

5. 冗余的set_dependency调用 #

该函数是tidymodels内部依赖注册函数，用户代码无需手动调用，直接删除即可。

修正后的完整代码 #

预处理Recipe #

data.split = initial_split(house.prices, prop=0.8)

train.data = training(data.split)
test.data = testing(data.split)

# Create recipe - 修正版
sales.rec = recipe(SalePrice ~., data = train.data) %>%
  step_log(SalePrice, LotArea, GrLivArea, TotRmsAbvGrd) %>%
  update_role(Id, new_role = "ID") %>%
  step_num2factor(MSSubClass, levels = as.character(unique(train.data$MSSubClass))) %>%
  step_unknown(PoolQC, Fence, MiscFeature, BsmtQual, BsmtCond, BsmtExposure, BsmtFinType1,
               FireplaceQu, GarageType, GarageQual, GarageCond, new_level = "None") %>%
  step_mutate(PorchArea = OpenPorchSF+EnclosedPorch+`3SsnPorch`+ScreenPorch) %>%
  step_mutate(garage.age = YrSold - GarageYrBlt,
         house.age = YrSold - YearBuilt,
         renovation.age = YrSold - YearRemodAdd
         ) %>%
  step_mutate(has.garage = !is.na(GarageType),
         has.basement = !is.na(BsmtExposure),
         has.pool = !is.na(PoolQC),
         is.new = house.age==0
         ) %>%
  step_cut(OverallQual, breaks = c(2.5, 6.5, 8.5)) %>%
  step_cut(OverallCond, breaks = c(2.5, 6.5, 8.5)) %>%
  step_mutate(house.age = log(1+house.age)) %>%
  step_rm(-c(SalePrice, house.age, renovation.age, 
             has.garage, has.basement, has.pool, is.new,
             MSZoning, LotArea, Alley, LotShape, Utilities, Neighborhood, OverallQual, OverallCond, 
             ExterCond, Foundation,
             BsmtFinSF1, Heating, HeatingQC, CentralAir, GrLivArea, FullBath, KitchenQual, TotRmsAbvGrd,
             PavedDrive, PorchArea, MiscVal,SaleCondition)) %>% 
  step_other(all_nominal_predictors(), all_factor(), all_string()) %>%
  step_string2factor(all_string_predictors()) %>%
  step_nzv(all_predictors()) %>%
  step_impute_median(all_numeric_predictors()) %>%
  step_unknown(all_factor_predictors()) %>%
  step_normalize(all_numeric_predictors())

模型与工作流 #

my.rf = rand_forest(mtry = tune(), trees = 2000, min_n=tune()) %>%
  set_engine("ranger") %>%
  set_mode("regression")

tree.grid = expand.grid(min_n=c(2,14,27,40), mtry = c(4, 8, 12))
folds = rsample::vfold_cv(train.data, v = 5)
metric = metric_set(rmse)

# Random forest - 修正版
my.rf.rec = sales.rec %>%
  step_dummy(all_nominal_predictors()) %>%
  step_zv(all_predictors())
my.rf.wflow = workflow() %>%
  add_model(my.rf) %>%
  add_recipe(my.rf.rec)

my.rf.res = my.rf.wflow %>%
  tune_grid(
  resamples = folds,
  metrics = metric,
  grid = tree.grid
)

内容的提问来源于stack exchange，提问作者JPR