Abstract by Devin Johnson
Summarizing Distributions of Latent Structures
In Bayesian analysis, our goal is not only to fit the model (e.g., sample from the posterior), but also to summarize the posterior distribution. Without an effective summary of the posterior, inference or application may be confusing or implausible. Common posterior summaries include means, medians, and variances, but these do not apply to distributions with more complex structure. We consider summarizing the distributions of latent structures such as clusterings, feature allocations, and networks. We present the sequentially-allocated latent structure optimization (SALSO) method to minimize an objective criterion in order to obtain a point estimate based on a collection of randomly sampled partitions. SALSO is a stochastic search method involving a series of micro optimizations. Several objective criteria can be used, including squared error loss, absolute error loss, and Binder loss.