Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..

Generalizing Weisfeiler-Lehman Kernels to Subgraphs

Authors: Dongkwan Kim, Alice Oh

ICLR 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In experiments on eight real-world and synthetic benchmarks, WLKS significantly outperforms leading approaches on five datasets while reducing training time, ranging from 0.01x to 0.25x compared to the state-of-the-art.
Researcher Affiliation Academia Dongkwan Kim & Alice Oh KAIST, Republic of Korea EMAIL, EMAIL
Pseudocode Yes Algorithm 1: 1-WL Algorithm Algorithm 2: WLSk Algorithm: 1-WL for subgraphs with their k-hop neighborhoods
Open Source Code Yes We make our code available for future research1. 1https://github.com/dongkwan-kim/WLKS
Open Datasets Yes We employ four real-world datasets (PPI-BP, HPO-Neuro, HPO-Metab, and EM-User) and four synthetic datasets (Density, Cut-Ratio, Coreness, and Component) introduced by Alsentzer et al. (2020).
Dataset Splits Yes Table 1: Statistics of real-world and synthetic datasets. ...Dataset splits 80/10/10 80/10/10 80/10/10 70/15/15 80/10/10 80/10/10 80/10/10 80/10/10
Hardware Specification Yes When measuring the complete training time, we run models of the best hyperparameters from each model s original code, including batch sizes and total epochs, using Intel(R) Xeon(R) CPU E5-2640 v4 and a single Ge Force GTX 1080 Ti (for deep GNNs).
Software Dependencies No All models are implemented with Py Torch (Paszke et al., 2019) and Py Torch Geometric (Fey & Lenssen, 2019). We use the implementation of Support Vector Machines (SVMs) in Scikit-learn (Pedregosa et al., 2011). Specific version numbers for these software dependencies are not provided.
Experiment Setup Yes We do a grid search of five hyperparameters: the number of iterations ({1, 2, 3, 4, 5}), whether to combine kernels of all iterations, whether to normalize histograms, L2 regularization ({23/100, 24/100, ..., 214/100}), and the coefficient α0({0.999, 0.99, 0.9, 0.5, 0.1, 0.01, 0.001}). When combining with kernels on continuous features (Equation 4), we tune αfeature from the space of {0.0001, 0.001, 0.01, 0.05, 0.1, 0.15, 0.2, 0.25} and set αstructure = 1/(1 + αfeature).