How to Solve Error in randomforest.default(m, y, …) : can’t have empty classes in y

The error message Error in randomforest.default(m, y, ...) : can't have empty classes in y in R typically occurs when you’re trying to build a Random Forest model using the randomForest package, and the response variable (y) contains classes with no instances or data points. This leads to an issue since Random Forest requires every class in the response variable to have at least one observation.

In this post, we’ll go through the common causes of this error and how to resolve it.

---

Example to Reproduce the Error

Let’s go through an example that will trigger this error.

# Load the necessary library
library(randomForest)

# Example dataset
data(iris)

# Modify the iris dataset to introduce the error
# We'll remove all rows where Species == "setosa"
iris_mod <- subset(iris, Species != "setosa")

# Attempt to train a Random Forest model on this modified dataset
rf_model <- randomForest(Species ~ ., data = iris_mod)

In the code above, we’ve removed all instances of the “setosa” class from the iris dataset. When you try to run the randomForest model on iris_mod, you’ll encounter the error:

Error in randomforest.default(m, y, ...) : can't have empty classes in y

The error occurs because the Species column no longer contains any “setosa” values, leaving one of the classes empty.

---

Solution

The error in the original example occurs because we removed all instances of the “setosa” class, leaving the Species variable with only two classes (“versicolor” and “virginica”). The Random Forest algorithm in R does not allow any class to be completely absent from the response variable. Here’s how to fix the issue:

1. Include All Classes or Use Subsetting Carefully: Instead of removing the entire class, ensure that your dataset includes at least one instance of each class. If you’re using a subset of the data and one class is missing, you either need to:
   • Modify the subset to include all classes, or
   • Train the model using a different approach, such as limiting the response variable to only the existing classes.
2. Recode or Remove Unnecessary Classes: If your analysis does not require the missing class, you can adjust the response variable to only focus on the classes that are present in your dataset.

Let’s demonstrate how to resolve the error in this particular case by either:

1. Using a subset of data that includes all classes, or
2. Recoding the response variable to exclude the missing class.

Option 1: Keep All Classes in the Data

# Use the full iris dataset to ensure all classes are present
rf_model <- randomForest(Species ~ ., data = iris)
rf_model

Output:

Call:
 randomForest(formula = Species ~ ., data = iris) 
               Type of random forest: classification
                     Number of trees: 500
No. of variables tried at each split: 2

        OOB estimate of  error rate: 4.67%
Confusion matrix:
           setosa versicolor virginica class.error
setosa         50          0         0        0.00
versicolor      0         47         3        0.06
virginica       0          4        46        0.08

Let’s break down what the results mean:

Call:

• The command randomForest(formula = Species ~ ., data = iris) shows the model is built to predict the Species variable using all other variables in the iris dataset.

Type of Random Forest: Classification:

• This indicates that the Random Forest model is solving a classification problem, meaning it’s predicting categories (in this case, different species of flowers).

Number of Trees: 500:

• The model has built 500 decision trees. More trees generally help stabilize predictions, but if performance doesn’t improve with more trees, you may not need as many for future runs.

No. of Variables Tried at Each Split: 2:

• At each node split in the decision trees, the model randomly selects 2 out of the 4 features in the iris dataset (Sepal.Length, Sepal.Width, Petal.Length, Petal.Width) to find the best split.
• This number is derived from the default behavior of Random Forest for classification problems, which typically uses sqrt(p) variables, where p is the number of predictors.

Out-of-Bag (OOB) Error Rate: 4.67%:

• The OOB error is an estimate of how well the model generalizes to unseen data. In your case, the model misclassifies approximately 4.67% of the OOB samples.
• This is a relatively low error rate, indicating that the model performs well overall.

Confusion Matrix:

The confusion matrix shows how well the model predicted the species of each flower based on the test data:

• Setosa:
  • Predicted correctly for all 50 instances.
  • Class error: 0.00 (Perfect classification of “setosa”).
• Versicolor:
  • The model predicted 47 instances correctly but misclassified 3 as “virginica.”
  • Class error: 6% (3 misclassifications out of 50).
• Virginica:
  • The model predicted 46 instances correctly but misclassified 4 as “versicolor.”
  • Class error: 8% (4 misclassifications out of 50).

Key Insights:

• Setosa: The model performs perfectly on the “setosa” class, with no misclassifications.
• Versicolor: The model has a few issues with distinguishing between “versicolor” and “virginica.” 3 of the “versicolor” flowers were misclassified as “virginica,” leading to a class error of 6%.
• Virginica: The class error for “virginica” is higher at 8%, with 4 instances being confused with “versicolor.” This suggests that the model has more difficulty distinguishing between these two species, likely due to their more similar feature distributions compared to “setosa.”

Option 2: Recode the Response Variable

If you are only interested in a subset of classes and are okay with excluding the “setosa” class, you can modify the response variable to only contain the relevant classes.

# Recode Species to exclude "setosa"
iris_mod$Species <- factor(iris_mod$Species)

# Now train the Random Forest model
rf_model <- randomForest(Species ~ ., data = iris_mod)

By explicitly refactoring the Species variable, you remove the empty class. This tells R that the response variable only contains the “versicolor” and “virginica” classes, allowing the Random Forest model to be trained on just these two classes without encountering the error.

Conclusion

The error Error in randomforest.default(m, y, ...) : can't have empty classes in y is triggered when one or more classes in the response variable have no data points. To solve it:

• Either ensure all classes are represented in the dataset.
• Or recode the response variable to exclude any empty classes.

By applying one of these solutions, you can avoid this issue and successfully train your Random Forest model.

Congratulations on reading to the end of this tutorial!

For further reading on random forest in R errors, go to the article:

How to Solve R Error: randomForest NA not permitted in predictors

 

Go to the online courses page on R to learn more about coding in R for data science and machine learning.

Have fun and happy researching!