Gaussian Mixture Models for User Segmentation

Description

Segments users with probabilistic mixture models that give each user a membership probability per segment, selecting the number of segments with information criteria.

Gaussian mixture models represent a population as a weighted mixture of K multivariate normal distributions, each a segment with its own mean and covariance. Unlike k-means, which assigns each user to one cluster and assumes spherical, equal-size clusters, GMMs give soft memberships and allow elongated, correlated and unequal clusters, which better reflects behavioural data where users blend patterns. Parameters are estimated with expectation-maximisation; the number of segments and covariance structure are chosen with BIC, and stability is checked by refitting on bootstrap samples. In psychology the equivalent is latent profile analysis. Rigour includes transforming skewed behavioural features, standardising, inspecting entropy (how cleanly users are classified), validating segments against held-out outcomes, and interpreting them with qualitative research.

Attributes

Loss type
Unmigrated
Effort
Weeks

Steps

  1. Fit GMM to behavioral feature matrix (sessions × features).
  2. Select K using BIC (Bayesian Information Criterion) — minimize BIC across K=2..10.
  3. Each user gets a probability vector over K clusters.
  4. Interpret cluster means as behavioral archetypes.
  5. Uncertainty: high-entropy probability vectors (spread across clusters) = users who don't fit any archetype — design for them separately.

Sources

  • Pattern Recognition and Machine Learning — Springer (Bishop)
  • Finite Mixture Models (Geoffrey McLachlan & David Peel)
  • Practical Statistics for Data Scientists (Bruce & Bruce)

Also known as

  • GMM segmentation
  • Latent profile analysis
  • Probabilistic clustering

Categories

  • Analytics & Behaviour
  • Computational & Machine Learning