t-SNE

A dimensionality reduction technique built purely to make high-dimensional data visually intuitive.

Key Points

Built specifically for visualization, not general-purpose dimensionality reduction for modeling.
Focuses on preserving local neighborhood structure — points that were close in high-dimensional space stay close in the 2D/3D plot.
Does not preserve global distances well — the distance between distant clusters in a t-SNE plot is not meaningful.
Has a key hyperparameter called perplexity, which roughly controls how many neighbors it balances local vs global structure against.
Results can vary between runs and are highly sensitive to perplexity — always visualize with multiple settings.

What it is

t-SNE (t-distributed Stochastic Neighbor Embedding) is a technique that reduces high-dimensional data down to 2 or 3 dimensions for visualization, specifically preserving how close points are to their neighbors rather than the global structure of the data.

Why it exists

PCA preserves overall variance but can fail to reveal cluster structure that only shows up in local neighborhoods, especially for non-linear data. t-SNE exists specifically to make clusters and neighborhood relationships in high-dimensional data — like word embeddings or image features — visually obvious in a 2D plot.

Intuition

Imagine trying to lay out a crumpled 3D friendship map flat on a table, where you only care about keeping close friends near each other — you're allowed to totally distort how far apart different friend groups end up from one another. t-SNE does exactly this: it obsesses over preserving 'who's near whom' locally and doesn't care about preserving the big-picture distances between separate groups.

When to use

Use to visually explore whether high-dimensional data (e.g. embeddings) has natural cluster structure.
Use for qualitative exploration and communication, like visualizing how a neural network separates classes internally.

When not to use

Do not use t-SNE output as input features for another model — it's designed for visualization, not for preserving usable structure downstream.
Do not interpret cluster sizes or distances between clusters in a t-SNE plot as meaningful — only local neighborhoods are preserved.
Do not run it once and trust the result — always try a few different perplexity values and random seeds.

Common mistakes

Reading too much into the distance between two separate clusters in a t-SNE plot — that distance carries no real meaning.
Using a single perplexity value and assuming the resulting plot is 'the' true structure of the data.
Running t-SNE directly on very high-dimensional raw data without a faster initial reduction step (e.g. PCA first) — it can be slow.

Notes

UMAP is a newer, often faster alternative to t-SNE that tends to better preserve some global structure while still excelling at local neighborhoods.

In practice

Common for visualizing high-dimensional embeddings (word2vec, image features) in 2D to visually inspect cluster separation before or after modeling.

Was this concept helpful?

Spot an error or gap? Suggest an edit →

Related concepts

principal-component-analysiscurse-of-dimensionality