umap-learn

UMAP-Learn

Overview

UMAP (Uniform Manifold Approximation and Projection) is a dimensionality reduction technique for visualization and general non-linear dimensionality reduction. Apply this skill for fast, scalable embeddings that preserve local and global structure, supervised learning, and clustering preprocessing.

Quick Start

Installation

Current stable release: umap-learn 0.5.12 (released April 2026). Requires Python 3.9+ and depends on scikit-learn>=1.6, numba, pynndescent, numpy, and scipy. Pin to a verified release:

uv pip install umap-learn==0.5.12

Basic Usage

UMAP follows scikit-learn conventions and can be used as a drop-in replacement for t-SNE or PCA.

import umap
from sklearn.preprocessing import StandardScaler

# Prepare data (standardization is essential)
scaled_data = StandardScaler().fit_transform(data)

# Method 1: Single step (fit and transform)
embedding = umap.UMAP().fit_transform(scaled_data)

# Method 2: Separate steps (for reusing trained model)
reducer = umap.UMAP(random_state=42)
reducer.fit(scaled_data)
embedding = reducer.embedding_  # Access the trained embedding

Preprocessing requirement: Match preprocessing to the metric. For numeric Euclidean-style metrics, scale features before fitting so high-variance columns do not dominate. For cosine, binary, precomputed-distance, or mixed-feature workflows, choose preprocessing that matches the metric instead of blindly standardizing every column.

Typical Workflow

import umap
import matplotlib.pyplot as plt
from sklearn.preprocessing import StandardScaler

# 1. Preprocess data
scaler = StandardScaler()
scaled_data = scaler.fit_transform(raw_data)

# 2. Create and fit UMAP
reducer = umap.UMAP(
    n_neighbors=15,
    min_dist=0.1,
    n_components=2,
    metric='euclidean',
    random_state=42
)
embedding = reducer.fit_transform(scaled_data)

# 3. Visualize
plt.scatter(embedding[:, 0], embedding[:, 1], c=labels, cmap='Spectral', s=5)
plt.colorbar()
plt.title('UMAP Embedding')
plt.show()

Parameter Tuning Guide

UMAP has four primary parameters that control the embedding behavior. Understanding these is crucial for effective usage.

n_neighbors (default: 15)

Purpose: Balances local versus global structure in the embedding.

How it works: Controls the size of the local neighborhood UMAP examines when learning manifold structure.

Effects by value:

Recommendation: Start with 15 and adjust based on results. Increase for more global structure, decrease for more local detail.

min_dist (default: 0.1)

Purpose: Controls how tightly points cluster in the low-dimensional space.

How it works: Sets the minimum distance apart that points are allowed to be in the output representation.

Effects by value:

Recommendation: Use 0.0 for clustering applications, 0.1-0.3 for visualization, 0.5+ for loose structure.

n_components (default: 2)

Purpose: Determines the dimensionality of the embedded output space.

Key feature: Unlike t-SNE, UMAP scales well in the embedding dimension, enabling use beyond visualization.

Common uses:

Recommendation: Use 2 for visualization, 5-10 for clustering, higher for ML pipelines.

metric (default: 'euclidean')

Purpose: Specifies how distance is calculated between input data points.

Supported metrics:

Recommendation: Use euclidean for numeric data, cosine for text/document vectors, hamming for binary data.

Parameter Tuning Example

# For visualization with emphasis on local structure
umap.UMAP(n_neighbors=15, min_dist=0.1, n_components=2, metric='euclidean')

# For clustering preprocessing
umap.UMAP(n_neighbors=30, min_dist=0.0, n_components=10, metric='euclidean')

# For document embeddings
umap.UMAP(n_neighbors=15, min_dist=0.1, n_components=2, metric='cosine')

# For preserving global structure
umap.UMAP(n_neighbors=100, min_dist=0.5, n_components=2, metric='euclidean')

Supervised and Semi-Supervised Dimension Reduction

UMAP supports incorporating label information to guide the embedding process, enabling class separation while preserving internal structure.

Supervised UMAP

Pass target labels via the y parameter when fitting:

# Supervised dimension reduction
embedding = umap.UMAP().fit_transform(data, y=labels)

Key benefits:

Semi-Supervised UMAP

For partial labels, mark unlabeled points with -1 following scikit-learn convention:

# Create semi-supervised labels
semi_labels = labels.copy()
semi_labels[unlabeled_indices] = -1

# Fit with partial labels
embedding = umap.UMAP().fit_transform(data, y=semi_labels)

When to use: When labeling is expensive or you have more data than labels available.

UMAP for Clustering

UMAP serves as effective preprocessing for density-based clustering algorithms like HDBSCAN, overcoming the curse of dimensionality.

Best Practices for Clustering

Key principle: Configure UMAP differently for clustering than for visualization.

Recommended parameters:

Clustering Workflow

Install HDBSCAN separately for density-based clustering:

uv pip install hdbscan

import umap
import hdbscan
from sklearn.preprocessing import StandardScaler

# 1. Preprocess data
scaled_data = StandardScaler().fit_transform(data)

# 2. UMAP with clustering-optimized parameters
reducer = umap.UMAP(
    n_neighbors=30,
    min_dist=0.0,
    n_components=10,  # Higher than 2 for better density preservation
    metric='euclidean',
    random_state=42
)
embedding = reducer.fit_transform(scaled_data)

# 3. Apply HDBSCAN clustering
clusterer = hdbscan.HDBSCAN(
    min_cluster_size=15,
    min_samples=5,
    metric='euclidean'
)
labels = clusterer.fit_predict(embedding)

# 4. Evaluate
from sklearn.metrics import adjusted_rand_score
score = adjusted_rand_score(true_labels, labels)
print(f"Adjusted Rand Score: {score:.3f}")
print(f"Number of clusters: {len(set(labels)) - (1 if -1 in labels else 0)}")
print(f"Noise points: {sum(labels == -1)}")

Visualization After Clustering

# Create 2D embedding for visualization (separate from clustering)
vis_reducer = umap.UMAP(n_neighbors=15, min_dist=0.1, n_components=2, random_state=42)
vis_embedding = vis_reducer.fit_transform(scaled_data)

# Plot with cluster labels
import matplotlib.pyplot as plt
plt.scatter(vis_embedding[:, 0], vis_embedding[:, 1], c=labels, cmap='Spectral', s=5)
plt.colorbar()
plt.title('UMAP Visualization with HDBSCAN Clusters')
plt.show()

Important caveat: UMAP does not completely preserve density and can create artificial cluster divisions. Always validate and explore resulting clusters.

Transforming New Data

UMAP enables preprocessing of new data through its transform() method, allowing trained models to project unseen data into the learned embedding space.

Basic Transform Usage

# Train on training data
trans = umap.UMAP(n_neighbors=15, random_state=42).fit(X_train)

# Transform test data
test_embedding = trans.transform(X_test)

Integration with Machine Learning Pipelines

from sklearn.svm import SVC
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import StandardScaler
import umap

# Split data
X_train, X_test, y_train, y_test = train_test_split(data, labels, test_size=0.2)

# Preprocess
scaler = StandardScaler()
X_train_scaled = scaler.fit_transform(X_train)
X_test_scaled = scaler.transform(X_test)

# Train UMAP
reducer = umap.UMAP(n_components=10, random_state=42)
X_train_embedded = reducer.fit_transform(X_train_scaled)
X_test_embedded = reducer.transform(X_test_scaled)

# Train classifier on embeddings
clf = SVC()
clf.fit(X_train_embedded, y_train)
accuracy = clf.score(X_test_embedded, y_test)
print(f"Test accuracy: {accuracy:.3f}")

Important Considerations

Data consistency: The transform method assumes the overall distribution in the higher-dimensional space is consistent between training and test data. When this assumption fails, consider using Parametric UMAP instead.

Performance: Transform operations are efficient (typically <1 second), though initial calls may be slower due to Numba JIT compilation.

Scikit-learn compatibility: UMAP follows standard sklearn conventions and works in pipelines. Recent 0.5.x releases also improved feature-name support and compatibility with current scikit-learn validation APIs:

from sklearn.pipeline import Pipeline

pipeline = Pipeline([
    ('scaler', StandardScaler()),
    ('umap', umap.UMAP(n_components=10)),
    ('classifier', SVC())
])

pipeline.fit(X_train, y_train)
predictions = pipeline.predict(X_test)
feature_names = pipeline.named_steps['umap'].get_feature_names_out()

Advanced Features

Parametric UMAP

Parametric UMAP replaces direct embedding optimization with a learned neural network mapping function.

Key differences from standard UMAP:

Installation:

uv pip install "umap-learn[parametric-umap]==0.5.12"
# Installs the TensorFlow-backed Parametric UMAP extra.

Basic usage:

from umap.parametric_umap import ParametricUMAP

# Default architecture (3-layer 100-neuron fully-connected network)
embedder = ParametricUMAP()
embedding = embedder.fit_transform(data)

# Transform new data efficiently
new_embedding = embedder.transform(new_data)

Custom architecture:

import tensorflow as tf

# Define custom encoder
encoder = tf.keras.Sequential([
    tf.keras.layers.InputLayer(shape=(input_dim,)),
    tf.keras.layers.Dense(128, activation='relu'),
    tf.keras.layers.Dense(64, activation='relu'),
    tf.keras.layers.Dense(2)  # Output dimension
])

embedder = ParametricUMAP(encoder=encoder, dims=(input_dim,))
embedding = embedder.fit_transform(data)

Persistence: Save Parametric UMAP with its built-in Keras-aware methods rather than plain pickle:

embedder.save("parametric_umap_model", exclude_raw_data=True)

from umap.parametric_umap import load_ParametricUMAP
loaded = load_ParametricUMAP("parametric_umap_model")
new_embedding = loaded.transform(new_data)

Recent 0.5.12 fixes include Parametric UMAP retraining stability improvements and metric-gradient fixes, so prefer the pinned current release for neural-network workflows.

When to use Parametric UMAP:

Inverse Transforms

Inverse transforms enable reconstruction of high-dimensional data from low-dimensional embeddings.

Basic usage:

reducer = umap.UMAP()
embedding = reducer.fit_transform(data)

# Reconstruct high-dimensional data from embedding coordinates
reconstructed = reducer.inverse_transform(embedding)

Important limitations:

Example: Exploring embedding space:

import numpy as np

# Create grid of points in embedding space
x = np.linspace(embedding[:, 0].min(), embedding[:, 0].max(), 10)
y = np.linspace(embedding[:, 1].min(), embedding[:, 1].max(), 10)
xx, yy = np.meshgrid(x, y)
grid_points = np.c_[xx.ravel(), yy.ravel()]

# Reconstruct samples from grid
reconstructed_samples = reducer.inverse_transform(grid_points)

AlignedUMAP

For analyzing temporal or related datasets (e.g., time-series experiments, batch data):

from umap import AlignedUMAP

# List of related datasets
datasets = [day1_data, day2_data, day3_data]

# Relations map matching sample indices between consecutive datasets.
relations = [
    {day1_idx: day2_idx for day1_idx, day2_idx in matched_day1_to_day2},
    {day2_idx: day3_idx for day2_idx, day3_idx in matched_day2_to_day3},
]

# Create aligned embeddings
mapper = AlignedUMAP().fit(datasets, relations=relations)
aligned_embeddings = mapper.embeddings_  # List of embeddings

When to use: Comparing embeddings across related datasets while maintaining consistent coordinate systems. relations is required for meaningful alignment; each dictionary describes how samples in one dataset correspond to samples in the next.

Reproducibility

To ensure reproducible results, always set the random_state parameter:

reducer = umap.UMAP(random_state=42)

UMAP uses stochastic optimization, so results will vary slightly between runs without a fixed random state.

Setting random_state prioritizes deterministic output. Leave it unset when throughput matters more than exact repeatability, because UMAP can use more parallelism without a fixed seed.

Common Issues and Solutions

Issue: Disconnected components or fragmented clusters

Issue: Clusters too spread out or not well separated

Issue: Poor clustering results

Issue: Transform results differ significantly from training

Issue: Slow performance on large datasets

Issue: NaN or inf values in input data

Issue: All points collapsed to single cluster

Issue: Imports resolve to a local file instead of the real package

Resources

Official documentation

references/

Contains detailed API documentation:

Load these references when detailed parameter information or advanced method usage is needed.