Online Chinese Restaurant Process

Summary

Liu et al. propose an online inference algorithm for the Chinese Restaurant Process. This is really two separate ideas, rolled into one. The first idea, which makes inference tractable, is to sample from each clustomer’s table posterior. The second idea, is to allow for retroactive corrections if supervised cluster labels are provided. The idea is to track, for each table, both the number of customers that were originally assigned to the table but later moved elsewhere, and the number of customers that were originally assigned elsewhere and later moved to the table. Together, these two counts (for each table) allow for estimating the mismatch between the model and the true class labels and correcting for future data.