Uniform Manifold Approximation and Projection (UMAP): Practical Intuitions and Demonstrations

#machinelearning #datascience #dataanalysis

This video is a recording of a lecture originally delivered for students of the SSC IMPROVE program through the University of Granada School of Technology and Telecommunications Engineering in January 2024.

Uniform Manifold Approximation and Projection (UMAP) is a non-linear (or manifold) learning technique for unsupervised dimensionality reduction. It can be used either as a preprocessing technique for other machine learning methods, or by itself as a method of data analysis and visualisation.

In this video, we go over the underlying principles related to how the algorithm calculates a lower-dimensional representation of the data, and practical considerations for building a UMAP model that is robust to small changes in meta-parameters. The two main meta-parameters discussed here are the number of nearest neighbours, and the minimum distance each point must be separated by in the lower dimensional representation.

Some calculations for the graph representation of the data were taken form ‪@statquest‬ , specifically the dissimilarity scores demonstrated in the "Mathematical Details" video. Please check out both of Joshua Starmer's two videos on UMAP, which are really excellent:

   • UMAP Dimension Reduction, Main Ideas!!!  

and

   • UMAP: Mathematical Details (clearly e...  

The mammoth data was taken from work done at the Smithsonian Institute and analysed originally by MNoichi. It will be linked following the channel verification.

You can follow along with the provided Jupyter Notebooks through the ssc-granada GitHub repository which will be linked following channel verification.