A Persistent Weisfeiler-Lehman Procedure for Graph Classification

Authors: Bastian Rieck, Christian Bock, Karsten Borgwardt

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

Reproducibility Variable Result LLM Response
Research Type Experimental In the following, we describe the practical performance of our methods on numerous graph classification benchmark data sets.
Researcher Affiliation Academia 1Department of Biosystems Science and Engineering, ETH Zurich, 4058 Basel, Switzerland.
Pseudocode Yes Algorithm 1 Persistent Subtree Feature Generation
Open Source Code Yes Please refer to our repository5 for the code and additional experiments. 5https://github.com/Borgwardt Lab/P-WL
Open Datasets Yes We use common graph benchmark data sets in our experiments, comprising graphs from chemoinformatics problems (Debnath et al. 1991), toxicology prediction (Helma et al. 2001), protein function/structure prediction (Borgwardt et al. 2005, Dobson & Doig 2003), carcinogenicity prediction (Wale et al. 2008), and social network analysis (Leskovec et al. 2005, Yanardag & Vishwanathan 2015).
Dataset Splits Yes We follow the standard setup for graph classification and perform a 10-fold cross-validation that we repeat 10 times, reporting the average and standard deviation for all runs. For hyperparameter tuning, we use an inner 5-fold cross-validation on each of the training splits to perform a grid search.
Hardware Specification No The paper does not provide specific details about the hardware (e.g., CPU, GPU models, or memory) used for running the experiments.
Software Dependencies No The paper states 'We implemented our methods in Python' but does not specify version numbers for Python or any other software dependencies or libraries.
Experiment Setup Yes For hyperparameter tuning, we use an inner 5-fold cross-validation on each of the training splits to perform a grid search. As for the hyperparameters, we choose p {1, 2} for P-WL and P-WL-C, and h {0, . . . , 10} for methods based on WL features, whereas we choose C {0.1, 1, 10} for training an SVM on the P-WL-D kernel values. Moreover, since we did not observe an effect in changing σ for P-WL-D, we leave σ = 1.0 fixed.