A concentration parameter, written α, controls how mixture weight is distributed across components in a stick-breaking process. Small α favors a few large clusters; large α spreads weight across many smaller clusters. section 8.12
The value of α changes the prior shape of the mixture:
- With α = 0.1, the process tends to create one dominant weight followed by weights close to zero. section 8.9
- With α = 1, break sizes are equally likely, producing a handful of comparatively substantial clusters. section 8.9
- With α = 10, many small breaks produce weights spread across numerous clusters. section 8.9
Thus, α expresses prior knowledge about how many components are plausible. After conditioning on data, Bayesian updating can assign large weights to supported clusters and negligible weights to the rest. section 8.9
Choosing a practical upper bound
An infinite stick-breaking process must be truncated for computation. A useful guide is an expected number of non-negligible components near α log N. For the Old Faithful data, α = 1 and K = 10 provide a loose bound; posterior weights identify two meaningful clusters and largely ignore the others. section 8.10
Figure 8.9 The effect of the concentration parameter 𝛼 of the beta distribution for the ratios 𝛽ᵢ on the stick-breaking weights 𝑤ᵢ
Compare how α changes beta-distributed break ratios and the resulting stick-breaking weights. section 8.9
