UMAP is a recent (2018) and increasingly popular technique for dimensionality reduction and visualization of high-dimensional data. UMAP is particularly well-suited for preserving the global structure of the data, including preserving the topological structure and the distances between points, which makes it an attractive alternative to popular dimensionality reduction methods such as t-SNE and PCA.

How Does UMAP Work?#

UMAP works by using a cost function that balances the preservation of local distances, global distances, and the preservation of the topological structure. This cost function is optimized using gradient descent. The result is a low-dimensional representation of the data that can be easily visualized and analyzed.

UMAP has been applied to a wide range of data types, including gene expression data, image data, and text data, and has been shown to produce visually appealing and interpretable results in many cases.


  1. UMAP is capable of preserving the global structure of the data, including topological structure and distances between points, making it a good choice for visualization and exploration.

  2. UMAP is computationally efficient and scalable, making it suitable for large datasets.

  3. UMAP can handle high-dimensional data, unlike some other dimensionality reduction methods that are limited to low-dimensional data.


  1. UMAP is a relatively new method and has not been as thoroughly tested and validated as some other dimensionality reduction methods.

  2. UMAP may not perform well on data with complex non-linear structure or on data with non-uniform density.

Example Code#

import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns
import umap

from sklearn.datasets import load_digits

digits = load_digits()
X =
y =

reducer = umap.UMAP()
X_reduced = reducer.fit_transform(X)

plt.scatter(X_reduced[:, 0], X_reduced[:, 1], c=y, cmap='viridis')


In conclusion, UMAP is a powerful and flexible tool for dimensionality reduction and visualization of high-dimensional data. While it is a relatively new method, it has already shown promising results and has the potential to become a widely used technique in the field of machine learning and data analysis.