使用mlr3tuning空间交叉验证调参时遇task_row_hash错误

使用mlr3tuning优化分类模型，尝试通过空间交叉验证（spcv_coords）对2个超参数调优时，执行test$train(task_sp)触发以下错误：

Error in if (!is.null(resamplings[[resampling]]$task_row_hash) && resamplings[[resampling]]$task_row_hash != :
missing value where TRUE/FALSE needed

报错运行日志 #

> test$train(task_sp)
INFO  [00:13:46.180] [bbotk] Starting to optimize 2 parameter(s) with '<OptimizerBatchGridSearch>' and '<TerminatorEvals> [n_evals=2, k=0]'
INFO  [00:13:46.207] [bbotk] Evaluating 9 configuration(s)
Error in if (!is.null(resamplings[[resampling]]$task_row_hash) && resamplings[[resampling]]$task_row_hash !=  : 
  missing value where TRUE/FALSE needed

原始代码 #

## Import data
df <- read.csv("C:/Users/Downloads/test.csv")
df[,c("presence")] <- as.factor(df[,c("presence")])
df[,c("habitat")] <- as.factor(df[,c("habitat")])
df[,c("species")] <- as.factor(df[,c("species")])
## summary(df)

## Create a task
task_sp <- mlr3spatial::as_task_classif_st(id = "A", x = df[,c("presence", "x", "y", "habitat")], target = "presence", positive = "1", coordinate_names = c("x", "y"), crs = "EPSG:4326", coords_as_features = FALSE)
task_sp$set_col_roles("presence", roles = c("target", "stratum"))

## Perform factor encoding
factor_encoding <- mlr3pipelines::po("encodeimpact", affect_columns = selector_cardinality_greater_than(10), id = "highCardinalityFactor") %>%
  mlr3pipelines::po("encode", method = "one-hot", affect_columns = selector_cardinality_greater_than(2), id = "lowCardinalityFactor") %>%
  mlr3pipelines::po("encode", method = "treatment", affect_columns = selector_type("factor"), id = "binaryFactor") %>%
  mlr3pipelines::po("imputeoor", affect_columns = selector_type("factor"), id = "outOfRangeFactor")
## print(po_factor_encoding)

## Perform tuning
test <- mlr3tuning::auto_tuner(tuner = mlr3tuning::tnr("grid_search", resolution = 5, batch_size = 10),
                               learner = mlr3::as_learner(factor_encoding %>% mlr3tuningspaces::lts(mlr3::lrn("classif.glmnet", predict_type = "prob", standardize = TRUE))),
                               resampling = mlr3::rsmp("spcv_coords", folds = 2),
                               measure = mlr3::msr("classif.prauc"),
                               terminator = mlr3tuning::trm("evals", n_evals = 2, k = 0))

test$train(task_sp)

错误核心原因是空间交叉验证采样器与分层角色设置冲突，同时因子编码管道存在逻辑重叠问题，以下是针对性修正：

1. 移除冗余的分层（stratum）角色设置 #

spcv_coords是基于空间坐标的交叉验证，不需要通过目标变量presence进行分层。将目标变量同时设为target和stratum会导致内部校验时出现空值判断错误，删除以下代码：

task_sp$set_col_roles("presence", roles = c("target", "stratum"))

2. 调整因子编码管道的逻辑顺序 #

原管道存在列选择重叠问题（如高基数因子会被后续的selector_type("factor")重复处理），调整顺序并明确各步骤的处理范围：

factor_encoding <- mlr3pipelines::po("imputeoor", affect_columns = selector_type("factor"), id = "outOfRangeFactor") %>%
  mlr3pipelines::po("encode", method = "treatment", affect_columns = selector_cardinality(2), id = "binaryFactor") %>%
  mlr3pipelines::po("encode", method = "one-hot", affect_columns = selector_cardinality_between(3,10), id = "lowCardinalityFactor") %>%
  mlr3pipelines::po("encodeimpact", affect_columns = selector_cardinality_greater_than(10), id = "highCardinalityFactor")

• 优先处理因子的缺失/异常值
• 明确二元因子（仅2个水平）用treatment编码
• 低基数多分类因子（3-10个水平）用one-hot编码
• 高基数因子（>10个水平）用encodeimpact编码

3. 修正调参器的参数匹配 #

原grid_search的resolution=5会生成25组参数组合，但终止器设为n_evals=2存在逻辑矛盾，建议根据测试需求调整：

• 快速测试：降低resolution并匹配n_evals
• 完整搜索：将终止器改为trm("none")或对应评估次数

修正后的完整代码 #

## Import data
df <- read.csv("C:/Users/Downloads/test.csv")
df$presence <- as.factor(df$presence)
df$habitat <- as.factor(df$habitat)
df$species <- as.factor(df$species)

## Create a task (移除冗余分层角色)
task_sp <- mlr3spatial::as_task_classif_st(
  id = "A", 
  x = df[,c("presence", "x", "y", "habitat")], 
  target = "presence", 
  positive = "1", 
  coordinate_names = c("x", "y"), 
  crs = "EPSG:4326", 
  coords_as_features = FALSE
)

## 调整后的因子编码管道
factor_encoding <- mlr3pipelines::po("imputeoor", affect_columns = selector_type("factor"), id = "outOfRangeFactor") %>%
  mlr3pipelines::po("encode", method = "treatment", affect_columns = selector_cardinality(2), id = "binaryFactor") %>%
  mlr3pipelines::po("encode", method = "one-hot", affect_columns = selector_cardinality_between(3,10), id = "lowCardinalityFactor") %>%
  mlr3pipelines::po("encodeimpact", affect_columns = selector_cardinality_greater_than(10), id = "highCardinalityFactor")

## 修正调参设置
test <- mlr3tuning::auto_tuner(
  tuner = mlr3tuning::tnr("grid_search", resolution = 2, batch_size = 5), # 降低分辨率适配测试
  learner = mlr3::as_learner(factor_encoding %>% mlr3tuningspaces::lts(mlr3::lrn("classif.glmnet", predict_type = "prob", standardize = TRUE))),
  resampling = mlr3::rsmp("spcv_coords", folds = 2),
  measure = mlr3::msr("classif.prauc"),
  terminator = mlr3tuning::trm("evals", n_evals = 4) # 匹配resolution生成的参数组合数
)

test$train(task_sp)

内容的提问来源于stack exchange，提问作者Sophie Père