Deepen the Understanding of WWL Graph Kernels

Original paper: Wasserstein Weisfeiler-Lehman Graph Kernels (NeurIPS 2019)

The codes are adapted from the accompanying code for the above paper. Please follow the README of that repository to install the dependencies.

Changes are made as follows:

1. Instead of manually downloading and decompressing the datasets, we use the DGL python library to fetch all the datasets that were used in the original paper.

2. main.py are modified to repeat cross-validation 10 times and report the average accuracy, similar to what the original paper did.

3. Add an argument --type, for users to choose whether run WWL discrete version (discrete), continuous version (continuous) or WWLsquare (both).

	MUTAG	PTC_MR	NCI1	PROTEINS	D&D	ENZYMES
WWL (from the paper)	87.27±1.50	66.31±1.21	85.75±0.25	74.28±0.56	79.69±0.50	59.13±0.80
WWL (our implementation)	87.81±1.46	65.69±1.32	85.66±0.15	74.89±0.68	79.38±0.39	58.28±1.07

	ENZYMES	PROTEINS(_full)	BZR	COX2	BZR_MD	COX2_MD
WWL (from the paper)	73.25±0.87	77.91±0.80	84.42±2.03	78.29±0.47	69.76±0.94	76.33±1.02
WWL (our implementation)	73.58±0.63	77.80±0.65	79.05±0.23	78.24±0.29	71.23±0.04	68.22±3.75
WWL^2	74.30±0.59	77.60±0.52	78.72±0.87	81.15±1.65	71.32±0.35	66.82±2.70