Uniform Manifold Approximation and Projection (UMAP)

Dimensionality Reduction in R

Matt Pickard

Owner, Pickard Predictives, LLC

PCA, t-SNE, and UMAP

PCA, t-SNE, and UMAP comparison

PCA, t-SNE, and UMAP

PCA, t-SNE, and UMAP comparison

PCA, t-SNE, and UMAP

PCA, t-SNE, and UMAP comparison

PCA, t-SNE, and UMAP

PCA, t-SNE, and UMAP comparison

PCA, t-SNE, and UMAP

PCA, t-SNE, and UMAP comparison

UMAP has similar hyperparameters that can be tuned.

UMAP plot

library(embed)


set.seed(1234)
umap_df <- recipe(Attrition ~ ., data = attrition_df) %>% 
  step_normalize(all_predictors()) %>% 
  step_umap(all_predictors(), num_comp = 2) %>% 
  prep() %>% 
  juice()


umap_df %>%  
  ggplot(aes(x = UMAP1, y = UMAP2, color = Attrition)) +
  geom_point(alpha = 0.7)

UMAP: employee attrition

UMAP plot of employee attrition

UMAP in tidymodels

Create recipe

umap_recipe <-  recipe(Attrition ~ ., data = train) %>% 
  step_normalize(all_predictors()) %>% 
  step_umap(all_predictors(), num_comp = 4)

Create model spec

umap_lr_model <- linear_reg()

UMAP in tidymodels

Create workflow

umap_lr_workflow <-  workflow() %>% 
  add_recipe(umap_recipe) %>% 
  add_model(umap_lr_model)

Fit the workflow

umap_lr_fit <- umap_lr_workflow %>% 
  fit(data = train)

UMAP in tidymodels

Evaluate the model

predict_umap_df <- test %>% 
  bind_cols(predict = predict(umap_lr_fit, test))

rmse(predict_umap_df, Attrition, .pred_class)

Let's practice!

Dimensionality Reduction in R