lmcinnes/umap

Features

• Manifold Learning Algorithms - Implements manifold learning algorithms to project high-dimensional data while preserving topological structure.
• Parametric Neural Mappings - Uses neural networks to learn a functional mapping from high-dimensional input to a low-dimensional embedding.
• Parametric Embedding Learning - Uses a neural network to learn the relationship between high-dimensional data and its low-dimensional projection.
• Joint Autoencoder Losses - Provides a joint autoencoder loss to optimize the neural network for simultaneous embedding and data reconstruction.
• Dimensionality Reduction - Projects high-dimensional data into a lower-dimensional space using non-linear techniques to facilitate visualization and analysis.
• Semi-Supervised - Incorporates partial label information by treating unlabelled points as noise, refining class separation when only some data is tagged.
• Supervised - Uses categorical labels during the embedding process to pull apart known classes while preserving data structure.
• Embedding Frameworks - Offers a framework for training parametric embeddings using neural networks to map high-dimensional spaces.
• Out-of-Sample Projections - Transforms new data points into a previously learned lower-dimensional embedding for downstream machine learning tasks.
• Parametric Manifold Learning - Trains neural networks to learn a mapping from high-dimensional spaces to low-dimensional embeddings.
• Supervised Embedding Learning - Uses categorical labels during the projection process to improve class separation and reveal data structures.
• Dimensionality Projection Plots - Generates dimensionality projection plots to visually identify clusters and trends in complex datasets.
• Mutual k-NN Graph Projections - Improves class separation in low-dimensional projections by using a mutual k-nearest neighbor graph to reduce distance concentration.
• Simplicial Set Intersection - Combines a fuzzy simplicial set with categorical label data to respect discrete metric distances.
• Fuzzy Simplicial Set Construction - Constructs membership strength data for local fuzzy simplicial sets using nearest neighbor indices and distances.
• Fuzzy Simplicial Set Constructions - Implements fuzzy simplicial set construction to build the weighted graphs necessary for manifold learning.
• Mutual K-Nearest Neighbor Graphs - Refines manifold connectivity using mutual k-nearest neighbor graphs to reduce hub effects and improve projection quality.
• Anomaly Detection Preprocessing - Reduces the complexity of high-dimensional datasets to improve the performance of density-based anomaly detection.
• Clustering Preprocessing - Reduces high-dimensional data to a lower-dimensional manifold to improve density-based clustering performance.
• Density-Based Anomaly Detectors - Identifies outliers in high-dimensional datasets by projecting data into a lower-dimensional space for efficient density-based detection.
• Density Regularization - Estimates local density of high-dimensional data and uses it as a regularizer for low-dimensional projections.
• Local Density Regularization - Adjusts the embedding process by estimating high-dimensional density to ensure relative spacing is preserved in the projection.
• Projection Plotting - Generates scatterplots of embeddings with automatic point-sizing and parameter watermarking for projection evaluation.
• Sparse Matrix Reduction - Provides direct dimensionality reduction for high-dimensional sparse matrices without requiring conversion to dense arrays.
• Embedding Alignment - Optimizes several low-dimensional projections simultaneously using shared-point constraints for consistent point locations.
• Landmark-Constrained Fine-Tuning - Fine-tunes learned embedding models to incorporate new information while preserving structure via landmark constraints.
• Incremental Embedding Updates - Appends new data slices to an existing aligned model by mapping shared points between slices.
• Landmark-Based Model Updating - Integrates new data into existing projections by anchoring the mapping to shared landmark points between data slices.
• Inverse Embedding Reconstruction - Trains a neural network to learn an inverse mapping from low-dimensional embeddings back to the original data space.
• Supervised Metric Learning - Trains a supervised embedding on labelled data and transforms unlabelled points into that space for feature engineering.
• Autoencoder Training Optimizations - Optimizes the encoder network using a joint loss function that combines dimensionality reduction goals with reconstruction accuracy.
• Time-Sequenced Alignment - Projects multiple time-sequenced data segments into a lower-dimensional space while preserving relative structure.
• Non-Euclidean Space Projections - Maps high-dimensional data into specialized output spaces such as spheres, toruses, or hyperboloids by specifying non-Euclidean distance metrics.
• Data Visualization Libraries - Provides a toolkit for rendering topological data visualizations, including scatterplots and graph connectivity views.
• Topological Graph Visualizations - Plots the weighted graph representing the manifold structure, featuring edge-bundling to reduce visual clutter.
• Non-Euclidean Metric Projections - Maps high-dimensional data into specialized geometric shapes like spheres or hyperboloids by utilizing non-Euclidean distance metrics.
• Interactive Visualization Toolkits - Provides zoomable, pannable interactive visualizations with data-driven tooltips for exploring individual points.
• Data Visualization - Dimensionality reduction for visualizing high-dimensional data.

Star history