Class separability and Overlap I have a dataset for five different classes with 40 features. This dataset is somehow imbalanced with 2 majority & 2 monirity classes, the other somehow average.
This is a classification task and were considering every class to have equal prediction priority (i.e. no priority as to correct prediction of one class is more important than the other.
Since models created on these dataset do not do well on both classes, I needed a way to analyse or visualize class separability in my dataset, so as to how separate or mixed/overlapped are these classes in my dataset.
But I want an idea about the right tool to explore these classes for to observer their separability. Multidimensional scaling or principal component analysis?
 A: A typical way to visualize in this way is t-distributed stochastic neighbor embedding (t-SNE). Briefly, t-SNE tries to create a low-dimension representation of high-dimensional data that respects, in some sense, the distance between points in the high-dimensional space. If the low dimension is $2$, you can plot the points in a scatterplot, giving a unique color to each class. t-SNE is available in the R package Rtsne and the Python package sklearn.
I have read some about how something called universal manifold approximation and projection (UMAP) is starting to be favored over t-SNE, though the math is harder, and the advantages UMAP has over t-SNE might not be worth it if you either do not understand the technique or have a customer who does not understand the technique.
Offhand, I know of two advantages UMAP has over t-SNE.

*

*UMAP seems to be faster.


*While t-SNE does not create a function that can be applied to new data, UMAP can be applied to new data.
I would consider the latter of those to be more important for someone interested in condensing many dimensions into fewer to have a low-dimension feature space for a regression or classification model that has some kind of out-of-sample validation (e.g., cross validation), rather than your objective of visualizing your data.
