An implementation of the Taxonomy Discovery for Personalized Recommendation paper on the MovieLens (small) dataset. The code is written in python. The original paper can be found here.
The algorithm alternates between latent factor updates and taxonomy path sampling. The latent factors are used to generate recommendations.
Path sampling uses the latent factors and the genres of the movies in a Gibbs sampling procedure to generate new paths for movies.
The initial taxonomy is generated randomly and then movies are assigned to nodes randomly. The Chinese restaurant process is not used.
As iterations progress, the taxonomy as well as the latent factors are updated using Gibbs sampling and gradient descent respectively.
Random initialization hit rate@10:
3%
Hit rate@10 after 10 iterations:
48%
A random sample of 10% of the movies are taken for the hit rate calculation.