A framework to evaluate the quality of dimensionality reduction procedures
DynamicViz, a Python package to visualize and analyze the stability in common dimension reduction techniques
Based on a quite simple idea, this tool allows you to explore how well (or bad!) your dimensionality reduction proceeded, thus helping you to detect distortions, make more robust interpretations )or perhaps diagnose pitfalls), compare and optimize the technique used, etc.
DynamicViz is a framework for generating dynamic visualizations of high-dimensional data that have been reduced to lower dimensions through dimensionality reduction (DR) methods such as principal component analysis (PCA) and t-distributed stochastic neighbor embedding (t-SNE), among others. These dynamic visualizations capture the sensitivity of DR visualizations to perturbations in the data and can be used to diagnose interpretative pitfalls, reinforce conclusions drawn from the analysis, or also to choose among different DR techniques.
As explained in the paper by Sun et al writing in Nat. Comput. Sci. 2022, the method begins by computing the DR on either the full dataset or bootstrapped samples of it. The reductions obtained from the bootstrap samples are then aligned to the reference. Then, the idea is straightforward…
