Sparse network learning with snlpy

Very large and sparse networks appear often in the wild and present unique algorithmic opportunities and challenges for the practitioner. While in theory, many operations on sparse networks can be implemented efficiently, this implementation is often missing in common libraries and is difficult to reliably implement oneself. In the python world, this is evident just looking for a library method to perform matrix multiplication on two sparse matrices.

The goal of snlpy is to provide the most efficient possible python implementations of useful machine learning algorithms on sparse networks. For now, the only algorithms provided are some common structural node embedding methods as well as an approximate ppr routine.

One major goal of snlpy is easy parallelization. This is achieved via the numba package. Wherever possible, algorithms execute in parallel and use all available cores. See the documentation for setting up your environment.

Implementations of the embedding algorithms already exist. These implementations are based on those found in karateclub. They have been modified to forego compatibility with networkx and to use numba for efficiency.

Installation

Simply clone the repo, navigate to the directory and

pip install snlpy/

Demo

There is a jupyter notebook outlining the basic features and usage of snlpy in the demo/ directory.