Description
Chooses the number of clusters or model complexity by balancing fit against parameter count, favouring the K that minimises the Bayesian information criterion.
BIC equals minus twice the log-likelihood plus the number of parameters times the log of sample size. Adding clusters always improves likelihood, but BIC's penalty grows with parameters, so the minimum indicates the best trade-off and approximates the model with the highest posterior probability. For Gaussian mixture segmentation, fitting K = 2 to 10 and comparing BIC across covariance structures gives a principled number of segments rather than one chosen to fit a slide. Differences in BIC above 10 are conventionally treated as very strong evidence. Rigour requires stable fits (multiple initialisations), checking the BIC curve rather than just its minimum, and judging whether extra clusters are substantively meaningful.
Attributes
- Discipline
- Statistics & Modelling
- Loss type
- Latent
- Goal
- Segment users
- Stage
- Post-launch
- Data
- Quantitative
- Effort
- Hours
Steps
- Fit candidate models for K = 2 to 10, each with multiple random initialisations (e.g. 10) and several covariance types.
- Compute BIC for each and plot BIC against K by covariance type.
- Identify the minimum; treat BIC differences below 2 as negligible, 2-6 as positive, 6-10 as strong, above 10 as very strong evidence.
- If the curve flattens, prefer the simpler model or use an elbow in BIC improvement.
- Confirm the chosen K with stability and validation indices and qualitative interpretability.
Sources
- Cluster Analysis for Researchers (Charles Romesburg)
- Pattern Recognition and Machine Learning (Christopher Bishop)
- Methods Atlas (Vivek Gopinath)
- Schwarz (1978) Estimating the dimension of a model
- Cluster Analysis for Researchers
- Pattern Recognition & ML (Bishop)
- Methods Atlas (Gopinath)
- The Inquiry Series, Ch 51 — Behavioral Clustering & Probabilistic Segmentation
Also known as
- Bayesian information criterion
- Schwarz criterion
Categories
- Analytics & Behaviour
- Statistics & Modelling