Probabilistic Models: temporal topic models and more

HDLSM: introduction

HDLSM stands for Hierarchical Dirichlet Latent Sequential Motifs. It uses Dirichlet Processes to improve over the PLSM model. Comparing to PLSM, the HDLSM model:

  • relaxes the constraint to have a fixed number of motifs K, and automatically finds the best number of motifs from the data,
  • replaces the table of starting time in each document p(z,ts|d) by an (infinite) mixture of motif occurrences, each occurrence being a pair (z,ts).

HDLSM Resources

More resources can be found in the article, the additional material, and in the web page on the presented CVPR poster.