3D Latent Space Analysis
Exploring the latent feature representations of datasets
One of the analysis techniques I used to search for explainable outliers was dimensionality reduction using algorithms such as UMAP [1]. A basic workflow is
where algorithmic choices are made for each step in the workflow. I partially simplified the algorithm by creating a jupyter notebook that combines the last three items into a single interface. Plotly made a nice interactive figure, which is shown below.
Visualizing the feature vectors in an interactive 3D embedding.
[1]UMAP