基于R语言随机森林算法的短信文本分类实现咨询

Hey there! Great job getting through data cleaning and creating your Document-Term Matrix (DTM)—that's half the battle for text classification tasks. Let’s walk through how to refine your random forest workflow, avoid common pitfalls, and get solid results with your SMS spam/ham dataset.

1. First: Ensure Consistent Train/Test Splitting #

Before jumping into model training, make sure your training and test sets are aligned correctly. A common mistake is creating separate DTMs for train and test data without enforcing the same vocabulary—this leads to mismatched feature dimensions that will break predictions later.

Here’s how to do it properly:

# Load required packages (if you haven't already)
library(tm)
library(randomForest)
library(caret)

# Split your original data into train and test sets (e.g., 70/30 split)
set.seed(123) # For reproducibility
train_idx <- createDataPartition(your_original_data$type, p = 0.7, list = FALSE)
train_data <- your_original_data[train_idx, ]
test_data <- your_original_data[-train_idx, ]

# Create corpus for train data, then DTM
train_corpus <- VCorpus(VectorSource(train_data$message))
# Apply your existing cleaning steps here (you mentioned you've done this)
sms_dtm_train <- DocumentTermMatrix(train_corpus, control = list(global = c(2, Inf)))

# **Critical**: Use the train DTM's vocabulary to build the test DTM
test_corpus <- VCorpus(VectorSource(test_data$message))
# Reapply cleaning steps to test corpus (same as train!)
sms_dtm_test <- DocumentTermMatrix(test_corpus, control = list(
  dictionary = Terms(sms_dtm_train), # Match train set's terms
  global = c(2, Inf)
))

2. Refine Your Random Forest Training Code #

Your initial code is on the right track, but let’s tweak it for better performance and robustness:

Adjust the ntree Parameter #

ntree=10 is way too small—random forests rely on many trees to average out noise and stabilize predictions. Aim for at least 500 trees (1000 is even better for larger datasets):

# Convert DTM to matrix (note: large DTMs may use lots of memory—see tip below)
train_matrix <- as.matrix(sms_dtm_train)

# Train the random forest classifier
set.seed(123) # Reproducibility
sms_classifier <- randomForest(
  x = train_matrix,
  y = train_data$type,
  ntree = 500, # Increase this for more stable results
  importance = TRUE, # Enable this to check feature importance later
  proximity = FALSE # Disable to save memory if not needed
)

Memory-Saving Tip for Large DTMs #

If your DTM is very large, as.matrix() can hog memory. Instead, convert it to a sparse matrix (using the Matrix package) which randomForest can handle:

library(Matrix)
train_sparse <- sparseMatrix(i = sms_dtm_train$i, j = sms_dtm_train$j, x = sms_dtm_train$v,
                             dims = sms_dtm_train$dim,
                             dimnames = sms_dtm_train$dimnames)

sms_classifier <- randomForest(
  x = train_sparse,
  y = train_data$type,
  ntree = 500,
  importance = TRUE
)

3. Evaluate Your Model #

Once your model is trained, test it on the held-out test set to measure performance:

# Prepare test matrix (match train format)
test_matrix <- as.matrix(sms_dtm_test)

# Generate predictions
predictions <- predict(sms_classifier, newdata = test_matrix)

# Create confusion matrix to assess accuracy, precision, recall
confusion_matrix <- table(Predicted = predictions, Actual = test_data$type)
print(confusion_matrix)

# Calculate metrics (using caret for nicer output)
confusionMatrix(predictions, test_data$type)

4. Optimize for Better Results #

Here are some ways to boost your model’s performance:

• Tune the mtry parameter: This controls how many features are randomly sampled for each tree. Use tuneRF() to find the optimal value:

  set.seed(123)
  tune_result <- tuneRF(train_matrix, train_data$type, ntreeTry = 500,
                        stepFactor = 1.5, improve = 0.01, trace = TRUE)
  

• Handle class imbalance: If spam messages are much rarer than ham, use the classwt parameter to weight minority class errors more heavily:

  sms_classifier <- randomForest(
    x = train_matrix,
    y = train_data$type,
    ntree = 500,
    classwt = c(ham = 1, spam = 5), # Adjust weights based on your class distribution
    importance = TRUE
  )
  

• Feature selection: Use variable importance to prune unhelpful terms. Plot top features with:

  varImpPlot(sms_classifier, n.var = 20) # Show top 20 most important terms
  

• Cross-validation: Use the caret package to run k-fold cross-validation for more reliable performance estimates:

  # Set up cross-validation control
  ctrl <- trainControl(method = "cv", number = 10, verboseIter = TRUE)
  
  # Train model with cross-validation
  rf_cv <- train(
    x = train_matrix,
    y = train_data$type,
    method = "rf",
    trControl = ctrl,
    ntree = 500,
    importance = TRUE
  )
  
  print(rf_cv)
  

5. Common Pitfalls to Avoid #

• Mismatched vocabularies: Always build the test DTM using the train set’s dictionary—never create separate DTMs from scratch for each split.
• Too few trees: ntree=10 will lead to unstable, high-variance predictions. Stick to 500+ trees.
• Ignoring class imbalance: If spam is a small subset, your model might default to predicting ham most of the time. Use class weights or resampling (e.g., SMOTE) to fix this.

内容的提问来源于stack exchange，提问作者Ayushi Misra