Julia and C++ implementations of Bayesian Probabilistic Matrix Factorization using Markov Chain Monte Carlo. BPMF is a recommender method that allows to predict for example movie ratings. The BPMF method is described here: http://www.cs.toronto.edu/~rsalakhu/papers/bpmf.pdf

The implementation in this repo is described in: "Distributed Matrix Factorization using Asynchrounous Communication", Tom Vander Aa, Imen Chakroun, Tom Haber, https://arxiv.org/pdf/1705.10633

Input is in the form of a sparse matrix of values (e.g. movie ratings) R. The outputs are two smaller matrices U and V such that U * V forms a prediction matrix.

Tutorial slides in tutorial_slides/

The single node OpenMP version of BPMF can be installed using Conda:

conda install -c vanderaa bpmf

CMake is used. The most interesting CMake options are:

• ENABLE_OPENMP: Enable OpenMP Support
• ENABLE_REDUCE: Reduce Mu/Lambda version
• BPMF_NUMLATENT: Number of latent dimensions

• BPMF_COMM: Communication library used, one of

  • GPI_COMM
  • MPI_ISEND_COMM
  • MPI_PUT_COMM
  • MPI_ALLREDUCE_COMM
  • MPI_BCAST_COMM
  • ARGO_COMM
  • NO_COMM

The C++ version takes these arguments:

Usage: bpmf -n <MTX> -p <MTX> [-o DIR/] [-i N] [-b N] [-krv] [-t N] [-m MTX,MTX] [-l MTX,MTX]

Paramaters:
  -n MTX: Training input data
  -p MTX: Test input data
  [-o DIR]: Output directory for model and predictions
  [-i N]: Number of total iterations
  [-b N]: Number of burnin iterations
  [-a F]: Noise precision (alpha)

  [-k]: Do not optimize item to node assignment
  [-r]: Redirect stdout to file
  [-v]: Output all samples
  [-t N]: Number of OpenMP threads to use.

  [-m MTX,MTX]: propagated posterior mu and Lambda matrices for U
  [-l MTX,MTX]: propagated posterior mu and Lambda matrices for V

Matrix Formats:
  *.mtx: Sparse or dense Matrix Market format
  *.sdm: Sparse binary double format
  *.ddm: Dense binary double format

The data/ directory contains preprocessed input data from the movielens database to predict movie ratings and from the chembl database to predict compound-on-protein activity.

The Julia version takes two command line arguments:

./bpmf <train_matrix.mtx> <test_matrix.mtx>

Matrices should be in the MatrixMarket format. Other options need to be changed in the source code itself. E.g. the number of iterations, the size of the features vectors, ...

The ArgoDSM version is being developed by Ioannis Anevlavis (Eta Scale AB).

For building, it is necessary to add the ArgoDSM installation folder to $PATH.

For running, you might need to change the size of the global address space (argo::init) based on the size of the globally allocated data structures.

For optimal performance, consider experimenting with a variety of optimizations that can be tuned through environment variables (for more info: https://etascale.github.io/argodsm/advanced.html).

Over the course of the last 5 years, this work has been supported by the EU H2020 FET-HPC projects EPEEC (contract #801051), ExCAPE (contract #671555) and EXA2CT (contract #610741), and the Flemish Exaptation project.