混合数据模糊聚类结果可视化技术问询（附R代码与数据样例）

Hi there! Let's build on the work you've already done with Gower distance and t-SNE to create clear, informative visualizations for your fuzzy clustering results. Here's a step-by-step guide with code examples tailored to your mixed categorical + numerical data:

1. Fix & Complete Preprocessing Steps #

I noticed a small typo in your Rtsne call (gower_dist1 should be gower_dist). Let's start by loading required packages and cleaning up the initial code for reproducibility:

# Load necessary packages
library(cluster)    # For daisy (Gower distance) and fanny (fuzzy clustering)
library(Rtsne)      # For t-SNE dimensionality reduction
library(ggplot2)    # For visualization
library(dplyr)      # For data manipulation
library(tidyr)      # For reshaping data

# Recreate your sample data
x <- data.frame(
  x1 = c("A", "B", "A", "B", "A", "B"),
  x2 = c("C", "C", "C", "D", "D", "C"),
  x3 = c(8.461373, 10.962334, 9.452127, 8.196687, 8.961367, 8.009029),
  x4 = c(27.62996, 27.22474, 27.57246, 27.29332, 26.72793, 27.97227),
  stringsAsFactors = FALSE
)

# Calculate Gower distance (ideal for mixed data types)
gower_dist <- daisy(x, metric = "gower")

# Run t-SNE to reduce to 2 dimensions
tsne_obj <- Rtsne(gower_dist, dims = 2, is_distance = TRUE)

2. Perform Fuzzy Clustering #

Since you're working with fuzzy clustering, we'll use the fanny() function from the cluster package. Adjust the k value (number of clusters) based on your domain knowledge or validation metrics:

# Run fuzzy clustering (example with k=2 clusters)
fuzzy_clust <- fanny(gower_dist, k = 2, metric = "gower")

# Extract fuzzy membership values (each row = 1 sample, columns = membership to each cluster)
membership_df <- as.data.frame(fuzzy_clust$membership)
colnames(membership_df) <- paste0("Cluster_", 1:ncol(membership_df))

# Add a hard label for the cluster with the highest membership (for quick reference)
membership_df$dominant_cluster <- factor(apply(membership_df, 1, which.max))

3. Combine Data for Plotting #

Merge the t-SNE coordinates, fuzzy memberships, and your original data into one dataframe for easy visualization:

# Create t-SNE coordinates dataframe
tsne_df <- data.frame(
  TSNE1 = tsne_obj$Y[,1],
  TSNE2 = tsne_obj$Y[,2]
)

# Combine all data into a single plot-ready dataframe
plot_data <- bind_cols(x, tsne_df, membership_df)

4. Visualization Options #

Option 1: Scatter Plot with Fuzzy Membership Transparency #

This plot uses color for the dominant cluster and transparency to show membership certainty (higher opacity = stronger membership to the dominant cluster):

ggplot(plot_data, aes(x = TSNE1, y = TSNE2)) +
  geom_point(
    aes(color = dominant_cluster, alpha = apply(select(plot_data, starts_with("Cluster_")), 1, max)),
    size = 3
  ) +
  scale_alpha(range = c(0.4, 1)) +
  labs(
    title = "t-SNE Visualization of Fuzzy Clustering Results",
    x = "t-SNE Dimension 1",
    y = "t-SNE Dimension 2",
    color = "Dominant Cluster",
    alpha = "Max Membership Value"
  ) +
  theme_minimal()

Option 2: Overlay Categorical Data #

Highlight your original categorical variables (x1, x2) by adding shape or secondary color to the plot:

ggplot(plot_data, aes(x = TSNE1, y = TSNE2)) +
  geom_point(
    aes(color = dominant_cluster, shape = x1, alpha = apply(select(plot_data, starts_with("Cluster_")), 1, max)),
    size = 3
  ) +
  scale_alpha(range = c(0.4, 1)) +
  labs(
    title = "t-SNE with Fuzzy Clusters + Categorical Variable x1",
    x = "t-SNE Dimension 1",
    y = "t-SNE Dimension 2",
    color = "Dominant Cluster",
    shape = "x1 Category",
    alpha = "Max Membership Value"
  ) +
  theme_minimal()

Option 3: Fuzzy Membership Heatmap #

Directly visualize membership distribution across clusters with a faceted heatmap:

# Reshape membership data for heatmap
membership_long <- plot_data %>%
  select(TSNE1, TSNE2, starts_with("Cluster_")) %>%
  pivot_longer(cols = starts_with("Cluster_"), names_to = "Cluster", values_to = "Membership")

ggplot(membership_long, aes(x = TSNE1, y = TSNE2)) +
  geom_tile(aes(fill = Membership)) +
  facet_wrap(~Cluster) +
  scale_fill_viridis_c(option = "plasma") +
  labs(
    title = "Fuzzy Membership Distribution Across Clusters",
    x = "t-SNE Dimension 1",
    y = "t-SNE Dimension 2",
    fill = "Membership Value"
  ) +
  theme_minimal()

Key Tips #

• Validate your cluster count k using metrics like the fuzzy silhouette coefficient (fuzzy_clust$silinfo$avg.width).
• Gower distance is critical here because it handles both categorical and numerical data without requiring normalization.
• Emphasize the fuzzy nature of your results (avoid treating them as hard clusters) – transparency and heatmaps are great ways to show membership uncertainty.

内容的提问来源于stack exchange，提问作者Jack shephard