fastwlk is a Python package that implements a fast version of the
Weisfeiler-Lehman kernel. It manages to outperform current state-of-the-art
implementations on sparse graphs by implementing a number of improvements
compared to vanilla implementations:
- It parallelizes the execution of Weisfeiler-Lehman hash computations since each graph's hash can be computed independently prior to computing the kernel.
- It parallelizes the computation of similarity of graphs in RKHS by computing batches of the inner products independently.
- When comparing graphs, lots of computations are spent processing positions/hashes that do not actually overlap between Weisfeiler-Lehman histograms. As such, we manually loop over the overlapping keys, outperforming numpy dot product-based implementations on collections of sparse graphs.
This implementation works best when graphs have relatively few connections compared to the number of possible connections and are reasonably dissimilar from one another. If you are not sure the graphs you are using are either sparse or dissimilar enough, try to benchmark this package with others out there using this script.
Running the benchmark script in examples/benchmark.py shows that for the
graphs in data/graphs.pkl, we get an approximately 80% speed improvement
over other implementations like grakel. The example dataset contains 2-nn
graphs extracted from 100 random proteins from the human proteome from the
AlphaFold EBI database.
To see how much faster this implementation is for your use case:
$ git clone git://github.com/pjhartout/fastwlk
$ poetry install
$ poetry run python examples/benchmark.pyYou will need to swap out the provided graphs.pkl with a pickled iterable of
graphs from the database you are interested in.