Transforming input features for t-SNE and UMAP

Question

Previously I was using PCA as my dimension reduction algorithm of choice, but have recently moved to using t-SNE and UMAP.

For PCA I would apply transformations to my input features to ensure they were approximately normal. For t-SNE I would apply the same transformations to features then the Rtsne function from the Rtsne package would apply PCA and input the PC's into the t-SNE model.

My question is what transformations or other dimension reduction techniques should I apply to my data prior to running UMAP (and for that matter is running with PC's in a t-SNE model the best way to analyse data)?

For my datasets many of the features represent ratios and I am not sure if this has any bearing on how my features should be treated.

Answer 1

For methods like t-SNE and UMAP it is the distance measure -- how you compare similarity between samples -- that really matters. While a default choice of euclidean distance is common, it is by no means necessary, and many other ways to measure distance may make more sense, particularly for different more specialised kinds of data. In terms of things to apply before applying UMAP or t-SNE, some kind of feature rescaling can make sense, but you can also think of this as a weighting on the components of the distance measure.