This repository provides a Python implementation - relying on the JAX library- of the Bernoulli Embeddings developed by Rudolph et al. (2016). Bernoulli Embeddings are a particular class of Exponential Family Embeddings where dense low-dimensional representations of items are constructued by relying on binary co-occurence patterns. This implementation underlies current work with Vasco M. Carvalho, Stephen Hansen, and Glenn Magerman that estimates these embeddings on 14 million firm-to-firm transactions for Belgium for the year 2014.
For a complete example on how to estimate these embeddings on a particular dataset check this blog post which uses movie ratings data to construct embedded representations of movies.