hsbm

Hierarchical stochastic block model (HSBM)

The code implements the algorithm discussed in Hierarchical Stochastic Block Model for Community Detection in Multiplex Networks.

The code now relies on the nett package for some of its functionality.

How to use:

• Install the nett package from the above link.
• Install the hsbm package from this repository by issuing the following command:

devtools::install_github("aaamini/hsbm", subdir = "hsbm_package")

• Run the benchmark.R in the root of the repository. The ouput would be something like this:

|method    | aggregate_nmi| slicewise_nmi| elapsed_time|
|:---------|-------------:|-------------:|------------:|
|HSBM      |        0.9766|        0.9804|        0.238|
|DP-SBM    |        0.1831|        0.8920|        1.197|
|SC-sliced |        0.0315|        0.2006|        0.098|
|SC-avg    |        0.0018|        0.0049|        0.035|
|SC-ba     |        0.0029|        0.0765|        0.086|
|SC-omni   |        0.0038|        0.0434|        0.199|
|PisCES    |        0.1016|        0.1408|        0.125|
|PisCES-sh |        0.0347|        0.1406|        0.094|

Besides HSBM, the following methods are implemented:

• DP-SBM refers to the algorithm that runs the Dirichlet Process SBM separately on each layer.
• SC-sliced runs the spectral clustering separately on each layer (i.e., slice).
• SC-avg run the spectral clustering on the average of the adjacency matrices from all layers.
• SC-bc is the biased-adjusted spectral clustering.
• SC-omni is the spectral clustering based on the omnibus embedding (with minor modification)
• PisCES The PisCES algorithm that solves an optimization problem that smooths out spectral projection matrices across time.
• PisCES-sh Our variation on the PisCES with shared k-means initialization. See the HSBM paper (updated version).

Changelog:

• 2/5/2021: The code was rewritten from scratch resulting in a much faster and more stable sampler. The code is optimized to work with sparse networks and runs ~ 100x faster than the old code (now moved to old_code/ folder). The sampler also mixes fast, on average in about 10 to 50 iterations. The new code is now provided in the R package hsbm available under the sub-folder hsbm-package.