t-distributed Stochastic Neighbor Embedding

A Culinary Example

t-SNE (t-distributed Stochastic Neighbor Embedding) is a tool that helps you visualize complex data.

When you have data with many features — like 10, 50, or even 100+ columns — it becomes impossible to see relationships between data points just by looking at a table. t-SNE helps by turning that complex, high-dimensional data into a simple 2D plot, where similar points end up close together and very different points are far apart.

To do this, t-SNE first measures how similar each point is to every other point based on their features. It calculates the distance between each pair of points — points that are close together are considered more similar. Then, it converts those distances into probabilities that reflect how likely each pair of points would be “neighbors.”

Finally, t-SNE tries to recreate those same neighbor relationships in a 2D space. The result is a map where you can visually explore the structure of your data, spot patterns, and often see natural groupings or clusters — even if the original data had dozens or hundreds of variables.

Overview

Suppose a researcher is studying how different beverages are perceived in terms of taste. They’ve asked participants to rate a variety of drinks — like soda, lemonade, energy drinks, milk-based beverages, and sparkling water — based on four characteristics:

• Sweetness

• Sourness

• Fizziness

• Creaminess

Each drink ends up with a score for each trait, turning every beverage into a point in 4-dimensional space.

The Problem:

With each drink described by four numbers, it’s hard to get a sense of how they all relate to one another.

• Are there groups of drinks that are experienced similarly?

• Which ones are completely unique?

Simply looking at the raw data doesn’t make patterns obvious.

This is where t-SNE comes in. t-SNE takes the 4D taste ratings and projects them into a 2D map, while trying to keep similar drinks close together and dissimilar ones farther apart — based on how people rated them.

t-SNE Output:

Drinks with similar flavor profiles are close together (like cola and sparkling water — both fizzy). On the 2D plot generated by t-SNE, you might notice:

• A tight cluster of carbonated drinks that are high in fizziness but low in creaminess.

• Another group of rich, smooth drinks that are creamy and not very fizzy.

• Some hybrids or outliers that don’t neatly fit either group — maybe a lightly fizzy, slightly creamy drink that stands alone.

You haven’t assigned these groups — t-SNE just arranged the data so these patterns are visually clear.

Why This Is Helpful:

• You get a visual overview of how items relate based on multiple features.

• It can guide decisions (e.g., comparisons, marketing, further analysis).

• And it’s especially useful when you’re working with many features and want a way to make sense of them.