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 approximate neighborhood structure, supervised learning, and clustering preprocessing.

Before interpreting an embedding, check finite coordinates, neighborhood retention, sensitivity to seeds/parameters, and domain evidence in the original space. UMAP axes, island areas, and gaps between clusters have no calibrated physical or probabilistic meaning. Keep predictive test data outside preprocessing and embedding fits.

Quick Start

Installation

Verified on 2026-10-01: umap-learn 0.5.12 is the current stable release (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 uses the scikit-learn estimator interface, but its geometry and supervised behavior differ from PCA and t-SNE. Core examples were exercised on small synthetic data with Python 3.13, scikit-learn 1.9.1, NumPy 2.5.3, and Numba 0.68.0. Parametric examples are source-checked illustrations, not executed TensorFlow tests. In snippets below, data is a finite samples-by-features array and labels must follow the same row order.

import umap
from sklearn.preprocessing import StandardScaler

# Scale continuous features when their units should carry equal weight
scaled_data = StandardScaler().fit_transform(data)

# Method 1: Single step (fit and transform)
embedding = umap.UMAP(random_state=42, n_jobs=1).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, decide whether scaling is scientifically appropriate; it changes the feature weighting. Fit preprocessing on training data only for predictive work. Use StandardScaler(with_mean=False) for sparse matrices to avoid densifying them. 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:

• Low values (2-5): Emphasizes fine local detail but may fragment data into disconnected components
• Medium values (15-20): Balanced view of both local structure and global relationships (recommended starting point)
• High values (50-200): Prioritizes broad topological structure at the expense of fine-grained details

Recommendation: Start with 15, keep it below the number of fitted samples, and compare several values. Larger neighborhoods do not make inter-cluster distances quantitatively reliable.

min_dist (default: 0.1)

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

How it works: Adjusts the attraction curve and typical packing in the embedding; it is not a hard lower bound on pairwise distances. Require 0 <= min_dist <= spread.

Effects by value:

• Low values (0.0-0.1): Creates clumped embeddings useful for clustering; reveals fine topological details
• High values (0.5-0.99): Encourages looser packing; does not certify preservation of global distances

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:

• 2-3 dimensions: Visualization
• 5-10 dimensions: Clustering preprocessing (may retain more useful structure than 2D; does not guarantee density preservation)
• 10-50 dimensions: Feature engineering for downstream ML models

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:

• Minkowski variants: euclidean, manhattan, chebyshev
• Spatial metrics: canberra, braycurtis, haversine
• Correlation metrics: cosine, correlation (good for text/document embeddings)
• Binary data metrics: hamming, jaccard, dice, russellrao, kulsinski, rogerstanimoto, sokalmichener, sokalsneath, yule
• Custom metrics: User-defined distance functions via Numba

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 emphasizing broader neighborhoods
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, encouraging class separation while retaining parts of the feature-neighborhood graph.

Supervised UMAP

Pass target labels via the y parameter when fitting:

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

Interpretation: Label-guided separation is part of the objective, not independent evidence of discovered classes. Tune using training folds and evaluate on held-out samples. To obtain an unsupervised fit, omit y; target_weight=0 with categorical labels still modifies graph edges in 0.5.12.

Semi-Supervised UMAP

For target_metric="categorical", encode known classes as nonnegative integers and unlabeled points as -1. This sentinel does not extend to arbitrary regression targets:

# 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 can help HDBSCAN on high-dimensional data, but reduction can create or erase clusters. Compare against clustering in the original or PCA space.

Best Practices for Clustering

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

Starting candidates, to validate on the actual data:

• n_neighbors: Compare 15, 30, and larger valid values; there is no universally correct neighborhood size
• min_dist: Set to 0.0 (pack points densely within clusters for clearer boundaries)
• n_components: Try 5-10 dimensions and compare cluster stability; standard UMAP does not preserve density

Clustering Workflow

Install HDBSCAN separately for density-based clustering:

uv pip install hdbscan==0.8.44

import umap
import hdbscan
from sklearn.preprocessing import StandardScaler

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

# 2. Candidate UMAP settings for clustering
reducer = umap.UMAP(
    n_neighbors=30,
    min_dist=0.0,
    n_components=10,  # Evaluate stability against other dimensions
    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
# Only when independently known labels are available; ARI here includes noise as -1.
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()

Validation: Report noise fraction and clusters found across seeds and parameter choices. Check cluster membership and domain evidence in the original feature space. A silhouette score computed only in the optimized embedding can be misleading. No cluster or all-noise output is a possible result, not a reason to tune until attractive islands appear.

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, random_state=42, stratify=labels
)

# 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: Validate transforms on held-out data representative of deployment. Inspect out-of-distribution inputs and neighborhood support before interpreting their locations; neither standard nor Parametric UMAP guarantees meaningful extrapolation under distribution shift. Retraining on representative data requires a new downstream validation.

Performance: Measure fit and transform with your sample size and metric; first calls include Numba compilation. There is no fixed latency guarantee. Standard transform preserves the trained coordinate system; update refits with added samples and can move existing points.

Scikit-learn compatibility: Pipeline.fit(X, y) forwards y to UMAP, so the following is a supervised UMAP pipeline. Keep the whole pipeline inside cross-validation; fitting the embedding once before cross-validation leaks information. The manual example above deliberately fits UMAP without labels:

from sklearn.pipeline import Pipeline

pipeline = Pipeline([
    ('scaler', StandardScaler()),
    ('umap', umap.UMAP(n_components=10, random_state=42, n_jobs=1)),
    ('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

densMAP

Use UMAP(densmap=True, output_dens=True) when local density is part of the question. fit_transform then returns (embedding, rad_orig, rad_emb): the radii are log-transformed local-radius estimates, not probabilities or raw density values. dens_frac=0.3 applies the density objective during the final 30% of optimization epochs, not to 30% of samples. densMAP supports Euclidean output and cannot transform unseen data or use inverse_transform in 0.5.12. See the reference for a runnable recipe.

Parametric UMAP

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

Key differences from standard UMAP:

• Uses TensorFlow/Keras to train encoder networks
• Enables efficient transformation of new data
• Supports reconstruction via decoder networks (inverse transform)
• Allows custom architectures (CNNs for images, RNNs for sequences)

Installation:

uv pip install "umap-learn[parametric-umap]==0.5.12"
# The released source also requires Keras >=3; verify the resolved TensorFlow/Keras pair.

Illustrative basic usage (TensorFlow runtime not exercised):

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:

from pathlib import Path
Path("parametric_umap_model").mkdir(exist_ok=True)
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)

Only load trusted model directories: loading reads a pickle. Verify transformed outputs after saving/loading. In 0.5.12, save writes parametric_model.keras while the loader checks parametric_model; do not assume full training-state restoration. exclude_raw_data=True is not a privacy/anonymization guarantee. The released source is authoritative where hosted examples differ.

When to use Parametric UMAP:

• Need efficient transformation of new data after training
• Require reconstruction capabilities (inverse transforms)
• Want to combine UMAP with autoencoders
• Working with complex data types (images, sequences) benefiting from specialized architectures

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:

• Approximate reconstruction, not a bijective inverse or evidence that a generated sample is plausible
• Unsupported for sparse training data, precomputed or gradient-free metrics, densMAP, and graph mode
• Computationally expensive; poor support outside the training embedding or between disconnected clusters

Explore inverse coordinates only where the training embedding supports them; rectangular grids often cross unsupported gaps. Check domain constraints on every generated reconstruction.

AlignedUMAP

For temporal or related datasets that need a shared coordinate system (time-series experiments, batches), use umap.AlignedUMAP().fit(datasets, relations=relations), where relations maps sample indices between consecutive datasets and is required for meaningful alignment. Parameters, methods, and a worked example are in references/api_reference.md under "AlignedUMAP Class" and "Usage Examples".

Reproducibility

For reproducibility in the same software/hardware environment, set random_state and retain the data order, preprocessing, package versions, and parameters:

reducer = umap.UMAP(random_state=42)

UMAP is stochastic; unseeded runs can differ substantially. Compare neighborhood or cluster stability rather than unaligned coordinates, which can rotate or reflect.

In standard UMAP, setting random_state forces n_jobs=1. transform_seed separately controls transform randomness. Exact cross-version or GPU/CPU equivalence is not promised. 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

• Solution: Inspect graph connected components, duplicates, metric support and outliers before increasing n_neighbors. Fully disconnected vertices can receive NaN coordinates; do not silently drop them from the analysis.

Issue: Clusters too spread out or not well separated

• Solution: Compare several min_dist values; visual separation is not a scientific validation criterion

Issue: Poor clustering results

• Solution: Compare neighborhood sizes/dimensions, noise fractions, multiple seeds, and a baseline without UMAP

Issue: Transform results differ significantly from training

• Solution: Check preprocessing and distribution shift; validate neighborhood support and downstream performance before reusing either UMAP variant.

Issue: Slow performance on large datasets

• Solution: Profile neighbor search versus optimization. low_memory=True reduces memory use, not necessarily runtime; consider PCA or sparse TruncatedSVD only when their information loss is acceptable

Issue: NaN or inf values in input data

• Solution: Impute or drop invalid rows before fitting. Current UMAP uses scikit-learn-style finite-value checks (ensure_all_finite) in fit() and update(), so clean numeric input is the safest default

Issue: All points collapsed to single cluster

• Solution: Check constant/duplicate features, the distance metric and preprocessing; compare multiple settings without assuming multiple true clusters exist

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

• Solution: Do not keep project files named umap.py, sklearn.py, hdbscan.py, or tensorflow.py beside notebooks or scripts. Those names can shadow installed packages and break or poison examples.

Resources

Official documentation

• UMAP user guide
• 0.5.12 release notes
• PyPI package (current stable: 0.5.12)
• GitHub repository

references/

Contains detailed API documentation:

• api_reference.md: Release-checked parameter contracts, return values, restrictions, and advanced recipes

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