Tasks · Established · Intermediate

Dimensionality Reduction

Compressing high-dimensional data into fewer dimensions while preserving as much meaningful structure as possible.

What Dimensionality Reduction is

High-dimensional data is sparse, slow and hard to visualise. Reduction techniques find a smaller coordinate system where the important variation survives.

How it works

PCA projects onto directions of maximum variance; t-SNE and UMAP preserve local neighbourhood structure for visualisation; autoencoders learn a compressed bottleneck representation.

Why it matters

It makes embeddings inspectable, speeds up downstream models and is the standard way to sanity-check whether a representation separates the classes you care about.

Common uses

→Visualising embedding spaces
→Noise reduction before modelling
→Compression for storage and retrieval

Strengths

✓Faster downstream training
✓Enables 2D visual inspection

Watch for

✓Distances in t-SNE and UMAP plots can mislead
✓Information is inevitably lost

Continue exploring

Embeddings Clustering Autoencoder