DNMF

Unofficial Python implementation of the Discrete Non-negative Matrix Factorization (DNMF) overlapping community detection algorithm


Paper

Ye, Fanghua, Chuan Chen, Zibin Zheng, Rong-Hua Li, and Jeffrey Xu Yu. 2019. “Discrete Overlapping Community Detection with Pseudo Supervision.” In 2019 IEEE International Conference on Data Mining (ICDM), 708–17. https://doi.org/10.1109/ICDM.2019.00081.


Requirements

  • python>=3.7.1
  • torch>=1.9.1

Quick start

  • To install the package run one of the two commands:

    • python -m pip install dnmf-python (installation from PyPI)
    • python setup.py install (compile from source, if cloned the repository)
  • To run the algorithm, load the graph adjacency matrix into a torch.FloatTensor (for ex. A), then call:

    from dnmf.DNMF import DNMF
    dnmf = DNMF()
    F = dnmf(A)
    
  • To run a quick test of the algorithm with an example graph, run python test.py from inside the src/dnmf/ directory


Config

The DNMF module supports the following hyperparameters as arguments:

  • alpha: tradeoff parameter for the U-subproblem
  • beta: tradeoff parameter for the F-subproblem
  • gamma: regularization parameter
  • k: desired number of overlapping communities
  • num_outer_iter: number of iterations for the outer loop (SDP iterations)
  • num_inner_iter: number of iterations for the inner loops (U and F subproblems)

Author

Andrej Janchevski

[email protected]

EPFL STI IEM LIONS

GitHub

View Github