UMAP

Description

Projects high-dimensional behavioural or text embeddings into two dimensions to visualise structure, clusters and outliers in user populations.

UMAP, listed in the treatise's borrowing matrix for machine learning, is a non-linear dimensionality reduction method that preserves local neighbourhood structure and more global structure than t-SNE. Applied to user feature vectors, session embeddings or embedded open-text responses, it produces maps where similar users or comments sit together, revealing clusters, continua and rare groups. For UX it is an exploratory instrument: it suggests segments and edge cases to investigate qualitatively. Rigour requires caution: distances between clusters and cluster sizes are not reliably meaningful, results depend on hyperparameters (n_neighbors, min_dist), and clusters should be confirmed with a separate method (for example HDBSCAN plus stability checks).

Attributes

Loss type
Active
Stage
Discovery
Effort
Days

Steps

  1. Build feature vectors per user or session, or embed text with a language model.
  2. Scale features and choose hyperparameters; run several settings.
  3. Project to 2D and colour by known attributes (plan, tenure, outcome).
  4. Cluster in the embedding or original space and test stability across seeds.
  5. Sample users from clusters and outliers for qualitative follow-up.
  6. Document caveats about interpreting distances.

Sources

  • Lost Methods in UX — A Complete Treatise (Gopinath)
  • McInnes L, Healy J & Melville J (2018) UMAP: Uniform manifold approximation and projection for dimension reduction. arXiv:1802.03426
  • Pattern Recognition and Machine Learning (Christopher Bishop)
  • Unsupervised Learning in Space and Time (Marc Leordeanu)
  • McInnes, Healy & Melville (2018) UMAP: Uniform Manifold Approximation and Projection for dimension reduction
  • Pattern Recognition & ML (Bishop)
  • Unsupervised Learning in Time & Space
  • The Inquiry Series, Ch 52 — Dimensionality Reduction — The Geometry of Behavioral Space

Also known as

  • Uniform Manifold Approximation and Projection
  • Behavioural embedding maps
  • UMAP for Behavioural Landscape Mapping

Categories

  • Analytics & Behaviour
  • Computational & Machine Learning