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..
Influence Functions for Edge Edits in Non-Convex Graph Neural Networks
Authors: Jaeseung Heo, Kyeongheung Yun, Seokwon Yoon, MoonJeong Park, Jungseul Ok, Dongwoo Kim
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments with real-world datasets demonstrate accurate influence predictions for different characteristics of GNNs. We further demonstrate that the influence function is versatile in applications such as graph rewiring and adversarial attacks. (Abstract); 4 Validation of influence function; 5 Applications |
| Researcher Affiliation | Academia | 1Graduate School of Artificial Intelligence 2Department of Computer Science & Engineering POSTECH, South Korea EMAIL |
| Pseudocode | No | The paper describes the Li SSA algorithm in Appendix D, but it is presented in paragraph form and mathematical equations, not as a structured pseudocode or algorithm block. |
| Open Source Code | Yes | Justification: We used public datasets such as Cora, and the code is provided in the supplementary material. (NeurIPS Paper Checklist, Section 5) |
| Open Datasets | Yes | We conduct experiments on five datasets: the citation graphs Cora, Citeseer, and Pubmed [29, 36], where the task is to predict each paper s research area based on citation relationships; and the Wikipedia graphs Chameleon and Squirrel [28], where the task is to estimate page traffic based on hyperlink relationships. |
| Dataset Splits | Yes | For the citation graphs, we follow the data splits provided by Yang et al. [36], and for the Wikipedia graphs, we use the splits from Pei et al. [27]. |
| Hardware Specification | Yes | All experiments are conducted using NVIDIA GeForce RTX 3090, NVIDIA RTX A5000, and NVIDIA RTX A6000 GPUs. (Appendix F) |
| Software Dependencies | No | For the adversarial attack experiments in Table 2, DICE [34] and PRBCD [14] are implemented by modifying the PyTorch-based Deep Robust library [24], while maintaining its default settings. This mentions a library but not its specific version or versions of other key software components like Python or PyTorch. |
| Experiment Setup | Yes | we tune the model and training hyperparameters of a vanilla GCN over the following search space: learning rates {0.1, 0.03, 0.01}, hidden dimensions {32, 64}, and weight decays {10 3, 10 4, 10 5, 10 6, 10 7}. Training is performed for 2000 epochs using the SGD optimizer. For influence function computation, we run the Li SSA algorithm for 10,000 iterations. The damping parameter λ is selected from {0.1, 0.01, 0.001, 0.0001}. (Appendix F) |