/sparsewl

Code for "Weisfeiler and Leman go sparse: Towards higher-order graph embeddings"

Primary LanguagePython

Weisfeiler and Leman go sparse

Code for "Weisfeiler and Leman go sparse: Towards scalable higher-order graph embeddings" (NeurIPS 2020).

Requirements

  • eigen3
  • numpy
  • pandas
  • scipy
  • sklearn
  • torch>=1.5
  • torch-geometric>=1.5
  • libsvm
  • pybind11

All results in the paper and the appendix can be reproduced by the following the steps below.

Reproducing the kernel experiments (precomputed Gram matrices) (Tables 1, 2a, 3a, 5, 6, 8, 9)

  • cd kernels
  • Download datasets from www.graphlearning.io, and place the unzipped folders into kernels/datasets
  • Download https://www.chrsmrrs.com/wl_goes_sparse_matrices/EXP.zip and https://www.chrsmrrs.com/wl_goes_sparse_matrices/EXPSPARSE.zip and unzip them into kernels/svm/GM
  • cd svm
  • Run python svm.py --dataset_dir ../datasets --gram_dir ../GM/EXP --k 1 --n_iters 5 --kernel WL --datasets ENZYMES

Reproducing the kernel experiments from scratch (Tables 1, 2a, 3a, 5, 6, 8, 9)

  • cd kernels
  • Download datasets from www.graphlearning.io, and place the unzipped folders into kernels/datasets
  • Run g++ gram.cpp src/*cpp -std=c++11 -o gram -O2
  • Create a directory for gram matrices mkdir ../GM/EXP -p
  • Run ./gram --dataset_dir ./datasets --gram_dir ../GM/EXP --k 1 --n_iters 5 --kernel WL --datasets ENZYMES (running times will be outputted on the screen, too)
  • cd svm
  • Run python svm.py --gram_dir ../GM/EXP --dataset_dir ../datasets --k 1 --n_iters 5 --kernel WL --datasets ENZYMES

New alternatives (faster)

  • cd kernels
  • Download datasets from www.graphlearning.io, and place the unzipped folders into kernels/datasets
  • Run python kernel.py --gram_dir ./GM/EXP --dataset_dir ./datasets --k 1 --n_iters 5 --kernel WL --datasets ENZYMES

Reproducing the neural baselines (Tables 1, 5)

  • cd neural baselines
  • Run python main_gnn.py

Reproducing the neural higher-order results (Table 2b, Figure 2abc, 3b, Table 7)

You first need to build the Python package:

  • cd neural_higher_order/preprocessing

  • You might need to adjust the path to pybind in preprocessing.cpp, then run

    • MaxOS: c++ -O3 -shared -std=c++11 -undefined dynamic_lookup python3 -m pybind11 --includes preprocessing.cpp src/*cpp -o ../preprocessingpython3-config --extension-suffix
    • Linux: c++ -O3 -shared -std=c++11 -fPIC python3 -m pybind11 --includes preprocessing.cpp src/*cpp -o ../preprocessingpython3-config --extension-suffix

  • Run the Python scripts in Alchemy, QM9, ZINC to reproduce the scores and running times

    • For example: cd Alchemy, python local_2_FULL.py to reproduce the scores for the \delta-2-LGNN on the Alchemy dataset