使用R语言mlr包绘制二分类目标偏依赖图遇问题求助

Hey there! Let's sort out why your partial dependence plots are only showing class labels instead of the probabilities you need. The issue almost certainly ties to how you've configured your classification task, learner, and partial dependence data generation—let's break down the fixes step by step.

Key Issues in Your Original Workflow #

From your code snippet, two critical pieces are likely missing:

• Your learner isn't configured to output probabilities (it's defaulting to class labels).
• The partial dependence data generation isn't set up to pull probability values instead of categorical predictions.

Corrected Full Workflow #

Here's the adjusted code with explanations for each fix:

library(mlr)
library(dplyr)
library(ranger)

# Prepare binary classification dataset
iris_bin <- iris %>% 
  filter(Species != "virginica") %>% 
  mutate(bin_target = ifelse(Species == "setosa", TRUE, FALSE)) %>% 
  select(-Species)

# 1. Define classification task with explicit positive class
task_bin <- makeClassifTask(
  data = iris_bin, 
  target = "bin_target", 
  positive = "TRUE"  # Critical: defines which class we want probabilities for
)

# 2. Configure learner to output probabilities (not just class labels)
lrn_ranger <- makeLearner(
  "classif.ranger", 
  predict.type = "prob",  # Core fix: tells the model to return probabilities
  num.trees = 100
)

# Train the model
model_bin <- train(lrn_ranger, task_bin)

# 3. Generate partial dependence data with probability extraction
pdp_data <- generatePartialDependenceData(
  model = model_bin, 
  task = task_bin, 
  features = "Sepal.Length",  # Replace with your target feature
  # Custom function to pull the positive class probability
  fun = function(model, newdata) {
    predict(model, newdata = newdata, type = "prob")$data[, "TRUE"]
  }
)

# Plot the PDP—this will now show probability values!
plotPartialDependence(pdp_data)

Why This Works #

Let's break down the critical changes:

• Explicit Positive Class: The positive = "TRUE" argument in makeClassifTask clarifies which category we're calculating probabilities for, eliminating ambiguity in binary classification.
• Probability-Focused Learner: Setting predict.type = "prob" in makeLearner ensures the ranger model outputs continuous probability values instead of discrete class labels.
• Custom Probability Extraction: The fun parameter in generatePartialDependenceData explicitly pulls the probability column for your positive class, ensuring the PDP data uses numerical values instead of categorical labels.

Quick Troubleshooting Check #

If you still run into issues:

• Verify that your bin_target column is a logical or factor (mlr handles these correctly for binary classification).
• Double-check that the feature name in features = "Sepal.Length" matches exactly with your dataset's column names.

内容的提问来源于stack exchange，提问作者stats-hb