Manifold learning

Description

A family of methods that assume high-dimensional user data lies on a lower-dimensional curved surface and recover its intrinsic coordinates.

Manifold learning starts from the idea that although users are described by hundreds of variables, real behaviour varies along a few underlying degrees of freedom, such as engagement and sophistication, arranged on a curved manifold. Methods such as Isomap (geodesic distances), locally linear embedding, Laplacian eigenmaps, diffusion maps, t-SNE and UMAP attempt to unfold it. For UX, the payoff is identifying the intrinsic dimensionality of behaviour and continuous trajectories, such as how users progress from novice to expert, rather than forcing discrete segments. Rigour involves estimating intrinsic dimension, comparing methods, checking neighbourhood preservation (trustworthiness and continuity scores) and validating interpretations with external data.

Attributes

Loss type
Latent
Effort
Weeks

Steps

  1. Construct and standardise the behavioural feature matrix, removing noise with PCA if dimensions are very high.
  2. Estimate intrinsic dimensionality (e.g. maximum likelihood estimator or PCA residual variance) to decide target dimensions.
  3. Apply two or more manifold methods (e.g. Isomap, diffusion maps, UMAP) with a range of neighbourhood sizes.
  4. Evaluate embeddings with trustworthiness and continuity metrics (values above 0.9 indicate good neighbourhood preservation).
  5. Interpret embedding coordinates by correlating them with original features and outcomes.
  6. Track user trajectories over time in the embedding to study progression or drift.

Sources

  • Pattern Recognition and Machine Learning (Christopher Bishop)
  • Unsupervised Learning in Space and Time (Marc Leordeanu)
  • Tenenbaum, de Silva & Langford (2000) A global geometric framework for nonlinear dimensionality 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

  • Non-linear dimensionality reduction
  • Nonlinear embedding

Categories

  • Analytics & Behaviour
  • Computational & Machine Learning