使用mlr3实现带样条的Cox glmnet模型调参时遇参数错误求助

尝试用mlr3复现带样条变换的cv.glmnet(family="cox")模型，训练AutoTuner时抛出错误：

Error in self$assert(xs, sanitize = TRUE) :
Assertion on 'xs' failed: Parameter 'alpha' not available. Did you mean 'spline_V1.applicator' / 'spline_V1.affect_columns' / 'spline_V2.applicator'?.

错误原因 #

当surv.glmnet学习器被嵌入到管道（GraphLearner）中后，其超参数不再是顶层的alpha和lambda，而是需要加上学习器在管道内的ID前缀。默认情况下，管道中的学习器ID为surv.glmnet，因此超参数的正确名称应为surv.glmnet.alpha和surv.glmnet.lambda。原代码中的搜索空间直接使用alpha和lambda，导致AutoTuner无法定位到这些属于管道子组件的参数。

修复步骤 #

1. 修正超参数搜索空间：为alpha和lambda添加管道内学习器的ID前缀
2. 可选优化：给管道中的学习器设置自定义ID，简化超参数名称

完整修复代码 #

require(mlr3)
require(mlr3proba)
require(mlr3learners)
require(mlr3tuning)
require(mlr3pipelines)
require(mlr3verse)
require(mlr3viz)
require(survival)
require(glmnet)
require(splines)

# 模拟回归数据集
set.seed(123)
n <- 100
p <- 3
X <- matrix(rnorm(n * p), nrow = n, ncol = p)
time <- rexp(n, rate = 1)
status <- sample(0:1, n, replace = TRUE)
df <- as.data.frame(X)
df$time <- time
df$status <- status

# 创建生存任务
task <- TaskSurv$new("survival_task", backend = df, time = "time", event ="status")

# 执行初始划分
initial_split <- rsmp("holdout")
initial_split$instantiate(task)

# 划分训练集与测试集
train_task <- task$clone()$filter(initial_split$train_set(1))  
test_task  <- task$clone()$filter(initial_split$test_set(1))

# 加载glmnet学习器
learner <- lrn("surv.glmnet")

# 修复：修改超参数搜索空间，添加学习器ID前缀
search_space <- ps(
  surv.glmnet.alpha  = p_dbl(lower = 0, upper = 1),   
  surv.glmnet.lambda = p_dbl(lower = 0.0001, upper = 0.1, logscale = TRUE)
)

# 定义样条变换函数
apply_splines <- function(x) {
  as.data.table(splines::ns(x, df = 3))  
}  

# 定义样条变换管道图
graph <- gunion(list(
  po("colapply", id = "spline_V1", applicator = apply_splines,        
      affect_columns = selector_name("V1")),
  po("colapply", id = "spline_V2", applicator = apply_splines,       
      affect_columns = selector_name("V2")),
  po("colapply", id = "spline_V3", applicator = apply_splines,
      affect_columns = selector_name("V3"))  )) %>% 
  po("featureunion") %>% 
  learner  # 默认分配ID为"surv.glmnet"

# 创建管道学习器 
pipeline <- GraphLearner$new(graph)

# 定义调参重采样策略
resampling <- rsmp("cv", folds = 5)

# 定义生存分析性能指标
measure <- msr("surv.cindex")

# 创建调参器
tuner <- tnr("grid_search", resolution = 5)

# 定义AutoTuner
at <- AutoTuner$new(
  learner = pipeline,
  resampling = resampling,
  measure = measure,
  search_space = search_space,
  terminator = trm("evals", n_evals = 20),
  tuner = tuner
)

# 在训练集上训练AutoTuner
at$train(train_task)

可选优化：自定义学习器ID #

若想简化超参数名称，可在管道中为学习器指定自定义ID：

# 修改管道定义，为学习器设置自定义ID
graph <- gunion(list(
  po("colapply", id = "spline_V1", applicator = apply_splines,        
      affect_columns = selector_name("V1")),
  po("colapply", id = "spline_V2", applicator = apply_splines,       
      affect_columns = selector_name("V2")),
  po("colapply", id = "spline_V3", applicator = apply_splines,
      affect_columns = selector_name("V3"))  )) %>% 
  po("featureunion") %>% 
  po("learner", learner = learner, id = "glmnet")  # 自定义ID

# 对应修改搜索空间
search_space <- ps(
  glmnet.alpha  = p_dbl(lower = 0, upper = 1),   
  glmnet.lambda = p_dbl(lower = 0.0001, upper = 0.1, logscale = TRUE)
)

内容的提问来源于stack exchange，提问作者atg