Using SimilaritySearch for ManifoldLearning

One of the main component of non-linear projection methods is the solution of large batches of k nearest neighbors queries. In this sense SimilaritySearch provides fast multithreaded algorithms using exact and approximated searching algorihtms.

This package provides two main structs UMAP, and ManifoldKnnIndex (to work with ManifoldLearning)


This package provides a pure Julia implementation of the Uniform Manifold Approximation and Projection dimension reduction algorithm

McInnes, L, Healy, J, Melville, J, UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction. ArXiV 1802.03426, 2018

The implementation in this package is based on the UMAP.jl package by Dillon Gene Daudert and collaborators. Forked and adapted to work with SimilaritySearch and take advantage of multithreading systems and different layout initializations. It can use any distance function from SimilaritySearch, Distances.jl, StringDistances.jl, or any distance function implemented by the user.

SimSearchManifoldLearning provides an implementation that partially supports the ManifoldLearning API using fit and predict, and similar arguments.

ManifoldLearning and ManifoldKnnIndex

We implemented the k nearest neighbor solver API of ManifoldLearning such that SimilaritySearch can be used. The use of SimilaritySearch gives multithreading k nearest neighbors solution, different distance functions and the different treadoffs between quality and speed for large datasets for the k nearest neighbors queries.


Some examples are already working in SimilaritySearchDemos.