Consider a Dirichlet process mixture model (DPM) with random precision parameter $\alpha$, inducing $K_n$ clusters over $n$ observations through its latent random partition. Our goal is to specify the prior distribution $p\left(\alpha\mid\boldsymbol{\eta}\right)$, including its fixed parameter vector $\boldsymbol{\eta}$, in a way that is meaningful. Existing approaches can be broadly categorised into three groups. Those in the first group rely on the linkage between $p\left(\alpha\mid\boldsymbol{\eta}\right)$ and $p\left(K_n\right)$ to draw conclusions on how to best choose $\boldsymbol{\eta}$ to reflect one's prior knowledge of $K_{n}$; we call them sample-size-dependent. Those in the second and third group consist instead of using quasi-degenerate or improper priors, respectively. In this article, we show how all three methods have limitations, especially for large $n$. We enrich the first group by working out and testing Jeffreys' prior in the context of the DPM framework, and by evaluating its behaviour. Then we propose an alternative methodology which does not depend on $K_n$ or on the size of the available sample, but rather on the relationship between the largest stick lengths in the stick-breaking construction of the DPM; and which reflects those prior beliefs in $p\left(\alpha\mid\boldsymbol{\eta}\right)$. We conclude with an example where existing sample-size-dependent approaches fail, while our sample-size-independent approach continues to be feasible.
翻译:暂无翻译