Machine Learning using R Dimensionality Reduction using UMAP

Readme.txt

This R script is a step by step demonstration on how to implement the Uniform Manifold Approximation and Projection (UMAP) using R.  Two R packages umap and uwot were demonstrated for both numeric data and mixed type data.   
A step by step video demonstration can be found at: https://youtu.be/Cf_pKrsDKi8?si=gHJp2bXPPzVa8M12 

R Script

#################################################################################
#                             Machine Learning using R   Dimensionality Reduction by using UMAP                                           #
#################################################################################

############## Brief Overview of UMAP (Uniform Manifold Approximation and Projection


### UMAP excels as a dimensionality reduction technique by preserving 
##  both local neighborhood structure and global data topology simultaneously.  

### Dual structure preservation: 
#      Maintains local clusters while capturing broad manifold relationships, 
#      unlike t-SNE (local-only) or PCA (linear global-only).  

### Computational efficiency: 
#      Scales to millions of points using approximate nearest neighbors, much faster 
#      than t-SNE for large datasets.  

###  Flexible embedding dimensions: 
#       Works for 2D visualization or higher-D preprocessing, 
#       unlike t-SNE's visualization limitation.  


###  Topology-aware: Mathematical foundation in Riemannian manifolds and fuzzy simplicial 
#        sets preserves holes, loops, and connectivity.  

###   Tunable parameters: `n_neighbors` balances local/global focus; 
#                         `min_dist` controls embedding tightness for interpretable control.  

###   UMAP serves visualization, clustering preprocessing, and feature engineering across 
#           genomics, NLP, and images. 


#### Generate a data set

set.seed(1222026)
n <- 1000  # observations
p <- 50    # dimensions

# Cluster 1: centered at origin
cluster1 <- matrix(rnorm(n * p, 0, 1), n, p)

# Cluster 2: shifted and scaled
cluster2 <- matrix(rnorm(n * p, 5, 0.5), n, p)

# Cluster 3: correlated features
cluster3 <- matrix(rnorm(n * p, -3, 1.5), n, p)
for (i in 1:n) {
  cluster3[i, ] <- cluster3[i, ] + seq(-2, 2, length.out = p)  # linear trend
}

# Combine and add labels
data <- rbind(cluster1, cluster2, cluster3)
labels <- factor(rep(1:3, each = n))
rownames(data) <- paste0("obs_", 1:nrow(data))
head(data)


#####  Load and run UMAP
###install.packages("umap")
set.seed(1222026)  # Remove random_state entirely

library(umap)
embedding <- umap(data, n_neighbors = 15, n_components = 2, 
                  min_dist = 0.1, metric = "euclidean")

# umap package returns matrix directly (not list)
str(embedding)  # 3000 x 2 matrix

# Create plot_df correctly
plot_df <- data.frame(X1 = embedding[,1], X2 = embedding[,2], label = labels)

ggplot(plot_df, aes(X1, X2, color = label)) +
  geom_point(size = 1, alpha = 0.7) +
  theme_minimal() +
  labs(x = "UMAP 1", y = "UMAP 2", title = "UMAP on Synthetic Clusters") +
  scale_color_brewer(type = "qual", palette = "Set1")

#### Load and run the R package uwot
#install.packages("uwot")
#library(uwot)

set.seed(1222026)  # for reproducibility

# Normalize data (center rows to mean 0 for cosine similarity)
data_norm <- scale(data, center = TRUE, scale = FALSE)

embedding_cosine <- uwot::umap(
  data_norm,
  n_neighbors  = 15,
  n_components = 2,
  min_dist     = 0.1,
  metric       = "cosine",
  n_threads    = 0
)


### Visualize Cosin distance
library(ggplot2)
plot_df_cosine <- data.frame(embedding_cosine, label = labels)
ggplot(plot_df_cosine, aes(X1, X2, color = label)) +
  geom_point(size = 1, alpha = 0.7) +
  theme_minimal() +
  labs(x = "UMAP 1 (Cosine)", y = "UMAP 2 (Cosine)", 
       title = "uwot UMAP with Cosine Metric") +
  scale_color_brewer(type = "qual", palette = "Set1")
   

#### Use uwot on mixed type data 
set.seed(1222026)

n <- 300

# Numeric block 1 (e.g. continuous measurements)
num1 <- matrix(rnorm(n * 5, 0, 1), ncol = 5)
colnames(num1) <- paste0("cont", 1:5)

# Numeric block 2 (e.g. counts, different scale)
num2 <- matrix(rpois(n * 3, lambda = 5), ncol = 3)
colnames(num2) <- paste0("count", 1:3)

# Categorical variables
cat1 <- factor(sample(letters[1:3], n, replace = TRUE))
cat2 <- factor(sample(c("low", "med", "high"), n, replace = TRUE))

mixed_df <- data.frame(num1, num2, cat1 = cat1, cat2 = cat2)
str(mixed_df)

 

set.seed(1222026)  # for reproducibility

emb_mixed <- uwot::umap(
  X = mixed_df,
  n_neighbors  = 15,
  n_components = 2,
  scale        = "Z",
  metric = list(
    euclidean   = c("cont1", "cont2", "cont3", "cont4", "cont5"),
    manhattan   = c("count1", "count2", "count3"),
    categorical = "cat1",
    categorical = "cat2"
  )
)

library(ggplot2)

plot_df <- data.frame(
  UMAP1 = emb_mixed[, 1],
  UMAP2 = emb_mixed[, 2],
  cat1  = mixed_df$cat1,
  cat2  = mixed_df$cat2
)

ggplot(plot_df, aes(UMAP1, UMAP2, color = cat1)) +
  geom_point(alpha = 0.7, size = 1.5) +
  theme_minimal() +
  labs(title = "uwot UMAP on mixed data", color = "cat1")

VideoDemo

A step by step video demonstration can be found at: https://youtu.be/Cf_pKrsDKi8?si=gHJp2bXPPzVa8M12