t-Distributed Stochastic Neighborhood Embedding (t-SNE)

Dimensionality Reduction in R

Matt Pickard

Owner, Pickard Predictives, LLC

t-SNE vs PCA table

Plotting PCA and t-SNE

PCA

PCA plot

Preserves global structure

t-SNE

t-SNE plot

Preserves local structure (keeps neighbors next to each other)

t-SNE hyperparameters

Perplexity - determines the number of nearest neighbors considered
Learning rate - rate the weights of the neural network are adjusted
Iterations - number of backpropogation iterations

t-SNE

t-SNE plot

t-SNE in R

library(Rtsne)


set.seed(1234)
tsne <- Rtsne(attrition_df %>% select(-Attrition))


tsne_df <- attrition_df %>% 
  bind_cols(tsne_x = tsne$Y[,1], tsne_y = tsne$Y[,2])


tsne_df %>% 
  ggplot(aes(x = tsne_x, y = tsne_y, color = Attrition)) +
  geom_point(alpha = 0.5)

t-SNE plot

Let's practice!

Dimensionality Reduction in R