Dimensionality Reduction

Dimensionality reduction is a family of unsupervised techniques that simplify data with many features into fewer dimensions while keeping as much of the important signal as possible. In SEO work it is most often used to shrink text embeddings, which can have hundreds of dimensions, so they can be clustered efficiently or plotted on a 2D chart.

How it works

Each column of data (or each number in an embedding) is a dimension. Reduction methods find a smaller set of new dimensions that preserve the relationships that matter:

• PCA (principal component analysis) finds the straight-line directions along which the data varies most. It is fast and predictable, and good for a first look or for compressing features.
• UMAP and t-SNE are non-linear methods that try to keep nearby points nearby. They produce much clearer cluster maps of embeddings; UMAP is the default inside BERTopic.
• Matrix factorisation methods such as LSA and NMF reduce word-document matrices into a small number of latent topics.

How it’s used in SEO and marketing

• Clustering pages or keywords: compress 384- or 768-dimension embeddings before running HDBSCAN or K-means, which improves speed and often cluster quality.
• Visualising a keyword universe or site as a 2D map where similar items sit together, useful for spotting gaps and overlaps.
• Entity analysis: MLforSEO’s query entity lab combines TF-IDF, K-means and dimensionality reduction to group entities by semantic relevance.
• Image clustering: pass images through a pre-trained network, reduce the feature vectors with PCA or UMAP, then cluster them.

Example

In MLforSEO’s guide to mapping keywords to topics, keywords are vectorised with TF-IDF, grouped with K-means, then reduced with PCA to two dimensions so each cluster can be plotted and sanity-checked by eye. With embeddings instead of TF-IDF, the same flow (embed, reduce with UMAP, cluster) is the core of BERTopic.

Limitations and things to watch

• Some information is always lost; check clusters in the original space or with real examples.
• Distances and cluster sizes on UMAP and t-SNE plots are not literal, so do not over-read gaps between groups.
• Non-linear methods are random unless you fix a seed, so results change between runs.
• PCA needs scaled features and only captures linear structure.

Learn it on MLforSEO

• Topic Modelling for SEO with BERTopic
• Different ways to Map Keywords to topics, with Supervised and Unsupervised ML approaches
• The Characteristics of a Machine Learning Problem
• Introduction to Machine Learning for SEO (course)