Homomorphism Expressivity of Spectral Invariant Graph Neural Networks

Authors: Jingchu Gai, Yiheng Du, Bohang Zhang, Haggai Maron, Liwei Wang

ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirically, a set of experiments on both synthetic and real-world tasks validate our theoretical results, showing that the homomorphism expressivity of spectral invariant GNNs well reflects their performance in down-stream tasks. [...] 4 EXPERIMENT In this section, we validate our theoretical findings through empirical experiments.
Researcher Affiliation Collaboration Jingchu Gai 1 Yiheng Du 1 Bohang Zhang1 Haggai Maron 2,3 Liwei Wang 1 1Peking University 2Technion 3NVIDIA Research EMAIL, EMAIL, EMAIL EMAIL, EMAIL
Pseudocode No The paper describes update rules for GNNs (e.g., in Definition 3.22, Definition E.1, Equation 7, Equation 8, Equation 9) but does not present them in explicitly labeled pseudocode or algorithm blocks. It provides mathematical formulations of the rules rather than structured, step-by-step algorithms formatted like code.
Open Source Code No The paper does not contain any explicit statement about releasing source code for the methodology described, nor does it provide any links to a code repository.
Open Datasets Yes Empirically, a set of experiments on both synthetic and real-world tasks validate our theoretical results, showing that the homomorphism expressivity of spectral invariant GNNs well reflects their performance in down-stream tasks. [...] Real-World Task We evaluate our GNN models on the ZINC-subset and ZINC-full dataset (Dwivedi et al., 2020).
Dataset Splits Yes For dataset setup and training parameters, we follow Zhang et al. (2024a).
Hardware Specification No The paper does not provide any specific hardware details such as GPU models, CPU types, or memory specifications used for running the experiments.
Software Dependencies No The paper does not specify any software dependencies or libraries with their version numbers that are needed to replicate the experiments.
Experiment Setup Yes For dataset setup and training parameters, we follow Zhang et al. (2024a). [...] The number of layers and hidden dimensions is set to match MPNN, such that all four models have roughly the same, and obey the 500K parameter budget in ZINC, as Zhang et al. (2024a) did.