T Sne and Umap for Visualization

The one idea: PCA compresses data by keeping the directions of maximum global variance — it's built to answer "what's the biggest spread in this dataset," not "which points are neighbors." t-SNE and UMAP are built for the second question: they optimize the low-dimensional layout specifically so that points close together in the original high-dimensional space stay close together on the 2D plot, even if that means distorting everything else about the geometry.

⚠️ The trap: A t-SNE plot with five tight, well-separated blobs is one of the most persuasive-looking charts in data science — and it is routinely over-read. Cluster distance, cluster size, and cluster density on a t-SNE or UMAP plot are not measurements of anything in the original data. Two clusters drawn 10 units apart are not "twice as different" as two clusters drawn 5 units apart; a fat cluster isn't necessarily more populous or more variable than a thin one. The only thing the plot promises to get right is which points are close and which are far — not by how much, and not the shape or size in between.

Why PCA falls short as a visualization tool

The PCA and LDA lesson covers PCA's mechanics: standardize, compute the covariance matrix, take its top eigenvectors (the directions of maximum variance), and project the data onto them. That's a great way to compress data while keeping as much total spread as possible. It is not the same goal as "make clusters visually separable."

Here's the mismatch, concretely. Imagine a dataset with three genuinely distinct, well-separated clusters — but one of them is much more spread out than…