Wasserstein Weisfeiler-Lehman Graph Kernels

Authors: Matteo Togninalli, Elisabetta Ghisu, Felipe Llinares-López, Bastian Rieck, Karsten Borgwardt

NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we analyse how the performance of WWL compares with state-of-the-art graph kernels. In particular, we empirically observe that WWL (1) is competitive with the best graph kernel for categorically labelled data, and (2) outperforms all the state-of-the-art graph kernels for attributed graphs.
Researcher Affiliation Academia 1DEPARTMENT OF BIOSYSTEMS SCIENCE AND ENGINEERING, ETH ZURICH, SWITZERLAND 2SIB SWISS INSTITUTE OF BIOINFORMATICS, SWITZERLAND
Pseudocode Yes Figure 1 illustrates the first two steps, and Algorithm 1 summarises the whole procedure.
Open Source Code Yes We leverage these resources and make our code publicly available1. 1https://github.com/Borgwardt Lab/WWL
Open Datasets Yes All the data sets have been downloaded from Kersting et al. [20].
Dataset Splits Yes As a classifier, we use an SVM (or a KSVM in the case of WWL) and 10-fold cross-validation, selecting the parameters on the training set only. We repeat each cross-validation split 10 times and report the average accuracy.
Hardware Specification Yes All our analyses were performed on a shared server running Ubuntu 14.04.5 LTS, with 4 CPUs (Intel Xeon E7-4860 v2 @ 2.60GHz) each with 12 cores and 24 threads, and 512 GB of RAM.
Software Dependencies No The paper mentions 'Available Python implementations can be used to compute the WL kernel [41] and the Wasserstein distance [13]' but does not provide specific version numbers for Python or these libraries. It lists Ubuntu 14.04.5 LTS as the operating system, but this is not an ancillary software component with a version number for replication.
Experiment Setup Yes As a classifier, we use an SVM (or a KSVM in the case of WWL) and 10-fold cross-validation, selecting the parameters on the training set only. We repeat each cross-validation split 10 times and report the average accuracy. We employ the same split for each evaluated method, thereby guaranteeing a fully comparable setup among all evaluated methods. Please refer to Appendix A.6 for details on the hyperparameter selection.