如何在R中绘制与roc.glmnet风格一致的Logistic回归及随机森林ROC曲线

I see exactly what's going on here—your current approach uses hard classification results (0/1 predicted labels) instead of predicted probabilities to calculate ROC curves, which is why you're stuck with those rough, two-point lines. Plus, we can standardize all three models' plots to match the polished style of glmnet's built-in ROC plot using the pROC package. Let's break this down step by step:

Why Your Current Code Fails #

• For Random Forest: You’re using as.numeric(rf.pred_SUM_db$predict) (the hard-classified 0/1 values) instead of the predicted probabilities of the positive class. This only gives you one threshold (0.5), resulting in a jagged line with just two data points.
• For Logistic Regression: If you’ve been using predict(type = "class") instead of type = "response", you’ll hit the same issue—only one threshold, no smooth curve to show model performance across all cutoff points.

Step 1: Fix Random Forest ROC Calculation #

Let’s adjust your Random Forest code to use positive-class probabilities instead of hard predictions:

# Load required packages
library(randomForest)
library(pROC)

# Keep your existing model fitting code
df_train_logit_rf_class$export_future <- as.factor(df_train_logit_rf_class$export_future)
df_test_rf_class$export_future <- as.factor(df_test_rf_class$export_future)

rf.fit_SUM_classification <- randomForest(
  formula = export_future ~ ., 
  data = df_train_logit_rf_class, 
  ntree = 500, 
  maxnodes = 100, 
  norm.votes = FALSE
)

# Extract POSITIVE CLASS probabilities (check colnames(rf.pred_SUM_db) to confirm column name)
rf.pred_SUM_db <- as.data.frame(predict(rf.fit_SUM_classification, df_test_rf_class, type = "prob"))
rf_pos_prob <- rf.pred_SUM_db[, "1"]  # Assumes "1" is your positive class (export_future=1)

# Calculate ROC using probabilities (not hard predictions!)
rf_roc <- roc(
  response = df_test_rf_class$export_future,  # pROC handles factor variables directly
  predictor = rf_pos_prob,
  levels = c("0", "1")  # Specify order: negative class first, positive second
)

Step 2: Fix Logistic Regression ROC Calculation #

For Logistic Regression, we’ll follow the same rule—use predicted probabilities instead of class labels:

# Fit Logistic Regression model
logit_fit <- glm(
  formula = export_future ~ ., 
  data = df_train_logit_rf_class, 
  family = binomial(link = "logit")
)

# Extract positive class probabilities
logit_pos_prob <- predict(logit_fit, newdata = df_test_rf_class, type = "response")

# Calculate ROC
logit_roc <- roc(
  response = df_test_rf_class$export_future,
  predictor = logit_pos_prob,
  levels = c("0", "1")
)

Step 3: Prepare LASSO ROC for Unified Plotting #

Your existing roc.glmnet call generates a valid ROC object, but we can convert it to a pROC object for consistent styling (or use it directly in plots):

library(glmnet)

# Use your existing LASSO fit
lasso_roc_obj <- roc.glmnet(lasso.fit_SUM, newx = x.train.loop, newy = y.train.loop)[[10]]
# Optional: Convert to pROC object for full styling control
lasso_roc <- roc(
  response = y.train.loop,
  predictor = predict(lasso.fit_SUM, newx = x.train.loop, s = lasso.fit_SUM$lambda[10], type = "response")
)

Step 4: Plot All Curves with Matching Aesthetics #

Now we’ll plot all three curves using pROC's plot() function, which replicates the clean, informative style of glmnet's default ROC plot (grid lines, AUC labels, smooth curves):

# Initialize plot with LASSO curve (base style matching glmnet)
plot(
  lasso_roc,
  col = "#1f77b4",  # Blue (matches glmnet's default)
  lwd = 2,
  main = "ROC Curves: LASSO vs. Logistic vs. Random Forest",
  grid = TRUE,  # Add grid lines like glmnet's plot
  print.auc = TRUE,  # Display AUC value
  print.auc.col = "#1f77b4",
  legacy.axes = TRUE  # Use 1-Specificity on x-axis (matches glmnet's output)
)

# Add Logistic Regression curve
lines(
  logit_roc,
  col = "#ff7f0e",  # Orange
  lwd = 2,
  print.auc = TRUE,
  print.auc.y = 0.4,  # Position AUC label below LASSO's
  print.auc.col = "#ff7f0e"
)

# Add Random Forest curve
lines(
  rf_roc,
  col = "#2ca02c",  # Green
  lwd = 2,
  print.auc = TRUE,
  print.auc.y = 0.3,  # Position AUC label below Logistic's
  print.auc.col = "#2ca02c"
)

# Add clean legend
legend(
  "bottomright",
  legend = c("LASSO", "Logistic Regression", "Random Forest"),
  col = c("#1f77b4", "#ff7f0e", "#2ca02c"),
  lwd = 2,
  bty = "n"  # Remove legend border for a polished look
)

Key Takeaways #

• Always use predicted probabilities (not hard classifications) for ROC curves—this gives you all possible threshold values, resulting in a smooth, informative curve that reflects model performance across cutoff points.
• The pROC package is flexible enough to match glmnet's aesthetic while letting you compare multiple models in a single plot.
• Double-check your positive class order in the roc() function to ensure the curve is calculated correctly.

内容的提问来源于stack exchange，提问作者Lusian