R语言中Kmeans预测工具包推荐及建模测试方案咨询

Hey Maria, sorry to hear the kmeans.predict() function from SwarmSVM is no longer available—let’s walk through how to split your dataframe into training/test sets and handle K-means clustering prediction in R, using both base R and some handy packages.

First, you’ll want to partition your data to train the model on a subset and assign clusters to unseen test data. Here are three reliable approaches:

Base R (No Extra Packages Needed) #

Use the built-in sample() function to randomly select indices for your training set:

# Assume your dataframe is named `df`
set.seed(123) # Fix random seed for reproducibility
train_size <- floor(0.7 * nrow(df)) # 70% for training, 30% for test
train_indices <- sample(seq_len(nrow(df)), size = train_size)

train_df <- df[train_indices, ]
test_df <- df[-train_indices, ]

Using the caret Package #

The caret package has a dedicated function for stratified splits (great if you want to preserve proportions of a categorical variable in your splits):

library(caret)
set.seed(123)

# If you have a categorical column (e.g., `df$category`), use it for stratification
train_indices <- createDataPartition(df$category, p = 0.7, list = FALSE)
# If no categorical column, use `1:nrow(df)` instead
# train_indices <- createDataPartition(1:nrow(df), p = 0.7, list = FALSE)

train_df <- df[train_indices, ]
test_df <- df[-train_indices, ]

Using the rsample Package #

rsample makes splitting data clean and readable with its intuitive split functions:

library(rsample)
set.seed(123)

data_split <- initial_split(df, prop = 0.7) # 70% training, 30% test
train_df <- training(data_split)
test_df <- testing(data_split)

Base R’s kmeans() function doesn’t include a built-in predict() method, but we can either implement this ourselves or use a package that adds this functionality.

Base R Implementation (Custom Predict Function) #

After training your K-means model on the training set, calculate distances from test samples to cluster centers and assign the closest cluster:

# Train K-means on the training set (e.g., 3 clusters)
set.seed(123)
kmeans_model <- kmeans(train_df, centers = 3)

# Custom predict function for K-means
predict_kmeans <- function(model, new_data) {
  # Calculate Euclidean distance from each test sample to every cluster center
  dist_matrix <- apply(new_data, 1, function(sample) {
    apply(model$centers, 1, function(center) {
      sqrt(sum((sample - center)^2))
    })
  })
  # Assign each sample to the cluster with the smallest distance
  cluster_assignments <- apply(dist_matrix, 2, which.min)
  return(cluster_assignments)
}

# Assign clusters to test data
test_cluster_labels <- predict_kmeans(kmeans_model, test_df)

Using the flexclust Package (Built-in Predict Method) #

The flexclust package provides a kcca() function (equivalent to K-means) that includes a native predict() method, making this process much simpler:

library(flexclust)
set.seed(123)

# Train K-means model with kcca
kcca_model <- kcca(train_df, k = 3, family = kccaFamily("kmeans"))

# Directly predict clusters for test data
test_cluster_labels <- predict(kcca_model, newdata = test_df)

• Always set a random seed (set.seed()) to ensure your splits and clustering results are reproducible.
• For K-means, make sure your data is scaled (using scale(df) if needed) before training, since the algorithm is sensitive to feature scales.

内容的提问来源于stack exchange，提问作者Maria Gold