基于caret包构建rpart模型及混淆矩阵相关技术咨询

Let's walk through fixing up your code first, then cover common tasks and issues related to your rpart model and confusion matrix work:

1. Fixing Your Initial Code #

First, I notice a couple of small issues in your current code snippet:

• You're setting fitControl twice (duplicate line) – you can remove one of those
• Your train() formula cuts off with Activit... – make sure to complete all predictor variables
• Also, don't forget to specify method = "rpart" in your train() call, since caret supports dozens of model types and needs clarity here

Here's the cleaned-up version of your code:

library(caret)
# Assuming you have data.table loaded for fread()
library(data.table)

data <- fread(sprintf("%s/%s", dataDir, dataFile), header = TRUE, stringsAsFactors = TRUE)
split <- createDataPartition(y = data$Audit, p = 0.7, list = FALSE)
trainSet <- data[split,]
testSet <- data[-split,]

# Single, clean fitControl definition
fitControl <- trainControl(method = "repeatedcv", number = 10, repeats = 10)

# Complete the formula and specify the rpart model method
fit <- train(Audit ~ Activity1 + Activity2 + Activity3, # Add all your predictors here
             data = trainSet,
             method = "rpart",
             trControl = fitControl)

2. Common Tasks for rpart Model & Confusion Matrix #

Getting Predictions & Generating Confusion Matrix #

Once your model is trained, you can generate predictions on the test set and create a confusion matrix to evaluate performance:

# Get predicted classes on the test set
testPred <- predict(fit, newdata = testSet)

# Generate the confusion matrix (compare predictions to actual values)
confMat <- confusionMatrix(testPred, testSet$Audit)

# Print the full details – includes accuracy, sensitivity, specificity, and class breakdown
print(confMat)

Tuning rpart Hyperparameters #

rpart's key hyperparameter is cp (complexity parameter), which controls tree pruning to avoid overfitting. You can tune it with a grid search:

# Define a range of cp values to test
tuneGrid <- expand.grid(cp = seq(0.01, 0.1, by = 0.01))

# Train the model with tuning
fitTuned <- train(Audit ~ Activity1 + Activity2 + Activity3,
                  data = trainSet,
                  method = "rpart",
                  trControl = fitControl,
                  tuneGrid = tuneGrid)

# Check which cp value performed best
print(fitTuned$bestTune)

Visualizing the Decision Tree #

To interpret your trained rpart model, plot the tree structure:

library(rpart.plot)
# Type 2 shows split labels below nodes; extra 104 adds class probabilities
rpart.plot(fit$finalModel, type = 2, extra = 104)

3. Troubleshooting Common Issues #

• Class Imbalance: If your Audit variable has heavily imbalanced classes, add classProbs = TRUE to trainControl and use metrics like ROC or sensitivity in the metric argument of train().
• Formula Errors: Double-check that all variable names in your formula match exactly what's in your dataset – typos (like your cut-off Activit...) will throw errors.
• Repeated CV Not Applying: Ensure you pass fitControl to train() via the trControl argument (it's an easy detail to miss!).

内容的提问来源于stack exchange，提问作者vik2426