Wasserstein Graph Distance Based on L1–Approximated Tree Edit Distance between Weisfeiler–Lehman Subtrees

Authors: Zhongxi Fang, Jianming Huang, Xun Su, Hiroyuki Kasai

AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct two types of experiments: metric validation and graph classification experiments. In the metric validation experiments, we demonstrate the effectiveness of the WWLS as a metric. In the graph classification experiments, we confirm the adaptability of the metric to graph classification, which represents one of its diverse applications.
Researcher Affiliation Academia Zhongxi Fang1, Jianming Huang1, Xun Su1, Hiroyuki Kasai1,2 1 Department of Computer Science and Communications Engineering, FSE Graduate School, Waseda University 2 Department of Communications and Computer Engineering, FSE School, Waseda University
Pseudocode Yes Algorithm 1: Enumerating all types of complete subtrees: DFSWL(G, h, v, d). Algorithm 2: Computing the WWLS distance.
Open Source Code Yes Source code: https://github.com/Fzx-oss/WWLS.
Open Datasets Yes We use TUD benchmark datasets, specifically selecting ten frequently used datasets, which can be grouped into two categories: (1) Bioinformatics datasets, including MUTAG, PTC-MR, COX2, ENZYMES, PROTEINS, NCI1, and BZR; and (2) Social network datasets, including IMDBB, IMDB-M, and COLLAB.
Dataset Splits Yes We employ a commonly used evaluation method for graph kernels (Morris et al. 2020), randomly splitting the data into a training set (90%) and a test set (10%), with a portion of the training set reserved for validation to tune the parameters.
Hardware Specification Yes The heavy processing parts of WWL and WWLS are written in C++, and we run programs on mac OS Monterey, Intel(R) Core(TM) i5-7360U CPU @ 2.30GHz.
Software Dependencies No The paper mentions C++ as the programming language and macOS Monterey as the operating system but does not provide specific version numbers for compilers, libraries, or other software dependencies.
Experiment Setup Yes The parameters of the WWLS are set as follows: we adjust the iteration number h within {2, 3} for the bioinformatics datasets and h = 1 for the social datasets due to their large node degrees; we adjust the parameter γ of Eq. (9) within {10 4, 10 3, . . . , 10 1}; and we adjust the regularization parameter C of SVM within {10 3, 10 2, . . . , 103}.