Tree Structure for the Categorical Wasserstein Weisfeiler-Lehman Graph Kernel
Authors: Keishi Sando, Tam Le, Hideitsu Hino
TMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Numerical experiments demonstrate that the performance of the proposed algorithm compares favorably with baseline kernels, while its computation is several orders of magnitude faster than the classic WWL graph kernel. |
| Researcher Affiliation | Academia | Keishi Sando EMAIL Department of Statistical Science The Graduate University for Advanced Studies Tam Le EMAIL The Institute of Statistical Mathematics Hideitsu Hino EMAIL The Institute of Statistical Mathematics / Waseda University |
| Pseudocode | Yes | Algorithm 1: Tree Structure for TWWL |
| Open Source Code | Yes | The experimental code is publicly available at https://github.com/Keishi-S/twwl. |
| Open Datasets | Yes | All datasets are downloaded from Morris et al. (2020). |
| Dataset Splits | Yes | In this experiment, we evaluate the performance of TWWL on a graph classification task using 10-fold cross-validation with a support vector machine classifier (Cortes & Vapnik, 1995). |
| Hardware Specification | Yes | All experiments are conducted on an Ubuntu 22.04 machine equipped with an Intel Xeon Gold 6354 CPU and 256GB of RAM. |
| Software Dependencies | No | The paper mentions that the method and WWL kernel are implemented in Julia and uses Python implementations for other methods, but does not provide specific version numbers for these software components or any other libraries. |
| Experiment Setup | Yes | For a regularization parameter C in the support vector classifier, we search over the range {10 3, 10 2, . . . , 103}. The number of iterations H for the WL algorithm is chosen from {1, . . . , 7}. The parameter λ of the WWL and TWWL is selected from {10 4, 10 3, . . . , 101}. For the Sinkhorn algorithm, we select the entropic regularization parameter from {0.01, 0.05, 0.1, 0.2, 0.5, 1, 10} and fix the maximum number of iterations to 1000, which is the default setting in the commonly used implementation (Flamary et al., 2021). In the second experiment, hyperparameter optimization is performed on the training data via grid search over the predetermined range of values. |