An employee survey with 200 questions is hard to compare across teams. Many answers move together: people who rate their manager highly tend to rate feedback and career growth highly too. Principal component analysis (PCA), the most common method, finds a handful of combined scores that capture most of the variation in all 200 answers. Analysts can then chart teams on three or four scores instead of 200, and a model trained on those scores has fewer inputs to overfit.
AI teams use the same idea to look at embeddings, which can have hundreds or thousands of numbers each. Methods such as t-SNE squash them onto a 2D map where similar items sit near each other. Those maps are for exploring, not measuring: the Distill guide shows that cluster sizes in a t-SNE plot carry no meaning and gaps between clusters may not either. Any reduction throws information away, so check whether the result still serves the task.
Methods such as PCA create new combined measurements, making them a form of feature extraction. Feature selection can also reduce dimensionality, but does it by keeping a subset of the original columns. An embedding model is itself a learned way of producing a representation; dimensionality reduction is a general technique you can apply to numeric data, including embeddings.
Sources
- scikit-learn: Principal component analysis (PCA) — Describes PCA as decomposing data into successive orthogonal components that explain a maximum amount of the variance; checked October 24, 2026.
- Wattenberg, Viégas and Johnson: How to Use t-SNE Effectively (Distill) — Shows with interactive examples that cluster sizes in a t-SNE plot mean nothing and distances between clusters might not mean anything; checked October 24, 2026.