Description
Compares hard k-means clustering with probabilistic Gaussian mixture models on the same behavioural data to choose a segmentation that reflects real structure and uncertainty.
K-means assigns each user to the nearest centroid, assuming spherical clusters of similar size, and gives no measure of uncertainty. Gaussian mixture models assume users are drawn from a mixture of multivariate normal distributions with their own means and covariances, yielding soft membership probabilities and support for elongated, overlapping or unequal clusters. Comparing them exposes whether segments are crisp or blended: if GMM posteriors are mostly above 0.9, segments are distinct; if many users sit near 0.5, hard personas will mislead. Rigour comes from standardising features, running multiple initialisations, comparing solutions with stability and validation indices, and checking that the chosen model's assumptions fit the data.
Attributes
- Discipline
- Computational & Machine Learning
- Loss type
- Latent
- Goal
- Segment users
- Stage
- Definition
- Data
- Quantitative
- Effort
- Days
Steps
- Build a standardised behavioural feature matrix and, if needed, reduce dimensions with PCA.
- Run k-means for K = 2-10 with at least 20 random initialisations per K.
- Fit GMMs for the same K range with full, diagonal and tied covariance options, selecting by BIC.
- Compare solutions with adjusted Rand index between methods and bootstrap stability (Jaccard above 0.75 indicates stable clusters).
- Inspect GMM posterior membership: report the share of users with maximum probability below 0.7 as ambiguous.
- Choose the method whose assumptions and stability fit, and document why.
Sources
- Cluster Analysis for Researchers (Charles Romesburg)
- Pattern Recognition and Machine Learning (Christopher Bishop)
- Methods Atlas (Vivek Gopinath)
- Cluster Analysis for Researchers
- Pattern Recognition & ML (Bishop)
- Methods Atlas (Gopinath)
- The Inquiry Series, Ch 51 — Behavioral Clustering & Probabilistic Segmentation
Also known as
- Hard versus soft clustering comparison
Categories
- Analytics & Behaviour
- Computational & Machine Learning