Abstract by Jeremy Meyer
Richard Warr, David Dahl
Attraction Indian Buffet Distribution
Latent feature models have been used to model many problems with a high number of predictor variables and few observations. The Indian Buffet process (IBP) is a popular Bayesian nonparametric prior for a latent feature allocation. This prior assumes exchangeability. In the presence of pairwise information, this assumption is untenable. There are cases where it is practical to use distances among observations for better results. We propose the attraction Indian Buffet distribution (AIBD), a generalization of the IBP that allows pairwise similarities to influence the sharing of features a priori. We explore the properties of the AIBD and conduct a simulation study that shows that the AIBD can leverage the pairwise distance information for improved feature allocation estimation.