core Estimated learning time: 5 h

5.15 Dimensionality reduction

You can compress and visualize high-dimensional data.

Before:03. Data Handling & AnalysisUnlocks:06. Deep Learning10. Production & MLOps13. Capstones, Portfolio & Interviews

Dimensionality reduction compresses many features into few: PCA for a faithful linear summary with variance-explained as its receipt, t-SNE and UMAP for visualisation. It sits in the unsupervised run because compression is structure-finding. The over-reading of t-SNE plots gets named directly — cluster sizes and between-cluster distances there are artefacts of perplexity settings, and treating the picture as geometry rather than suggestion is how false stories start.

Work through these

  • PCA: derivation, variance explained, whitening

    Finding the directions of greatest variation and keeping the strongest few, with a clear statement of how much information was kept. It is the workhorse and it is linear, which is both its strength and its limit.

  • t-SNE: perplexity and how to not over-read it

    A nonlinear method for visualisation, with one setting that changes the picture substantially and results that must not be over-interpreted. Distances between well-separated groups in the output mean very little.

  • UMAP and its trade-offs

    A more recent alternative that is faster and preserves more of the large-scale structure, with its own settings and its own caveats. It has largely become the default for visualisation.

  • Feature selection vs. feature extraction

    Choosing a subset of existing features and constructing new combined ones are different operations with different consequences for interpretability. Knowing which you need comes before choosing a method.

Sign in to keep your progress.

Free resources

Links last checked 29 Aug 2026.

Stuck here?

Ask a mentor. A real person answers, and they can see exactly which topic you're on. Usually within a couple of working days.

Checking your session…

Topics shown in module order.