On Explaining Equivariant Graph Networks via Improved Relevance Propagation

Authors: Hongyi Ling, Haiyang Yu, Zhimeng Jiang, Na Zou, Shuiwang Ji

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

Reproducibility Variable Result LLM Response
Research Type Experimental Through experiments on both synthetic and real-world datasets, our method demonstrates its capability to identify critical geometric structures and outperform alternative baselines.
Researcher Affiliation Academia 1Department of Computer Science and Engineering, Texas A&M University, Texas, USA 2Department of Industrial Engineering, University of Houston, Texas, USA.
Pseudocode No The paper describes the methodology using mathematical formulations and descriptive text, but does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks.
Open Source Code Yes Our code has been released as part of the AIRS library (https://github.com/divelab/AIRS/).
Open Datasets Yes In addition to synthetic datasets containing perfect 3D geometric shapes, we evaluate our method on three real-world datasets, including the Structural Classification of Proteins (SCOP), Bio Li P, and Actstrack. The SCOP database (Murzin et al., 1995; Andreeva et al., 2007; Chandonia et al., 2019) is a predominantly manually curated classification of protein structural domains... Bio Li P (Yang et al., 2012; Zhang et al., 2024) is a semi-manually curated database dedicated to high-quality ligand-protein interactions... Acts Track (Miao et al., 2023) is a particle tracking simulation dataset in high-energy physics.
Dataset Splits No The paper mentions using 'training and validation datasets' for SCOP by referencing external papers (Hou et al., 2018; Hermosilla et al., 2020), and mentions a 'test dataset' for Bio Li P, but does not provide explicit percentages, sample counts, or detailed splitting methodologies for all datasets, particularly the synthetic Shapes and Spiral Noise datasets.
Hardware Specification Yes We use NVIDIA RTX A6000 GPUs for all our experiments.
Software Dependencies No The paper does not provide specific version numbers for any software dependencies or libraries used in the experiments.
Experiment Setup Yes Table 4 provides details on the number of layers, the number of hidden equivariant features, and the highest order of equivariant feature lmax in the TFN.