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..
Geometry-Aware Edge Pooling for Graph Neural Networks
Authors: Katharina Limbeck, Lydia Mezrag, Guy Wolf, Bastian Rieck
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirical results demonstrate that our methods (i) achieve top performance compared to alternative pooling layers across a range of diverse graph classification tasks, (ii) preserve key spectral properties of the input graphs, and (iii) retain high accuracy across varying pooling ratios. 4 Experimental Results Across our experiments, we address four key tasks, namely (i) graph classification performance, (ii) graph structure preservation during pooling, (iii) performance across varying pooling ratios, and (iv) performance at graph property regression. |
| Researcher Affiliation | Academia | 1Helmholtz Munich 2Technical University of Munich 3Université de Montréal 4Mila Quebec AI Institute 5Université de Fribourg |
| Pseudocode | Yes | We provide a pseudocode implementation of our algorithm in Appendix C.4. ... Algorithm 1 Graph Pooling Methods: Spread Edge Pool and Mag Edge Pool |
| Open Source Code | Yes | Our methods are available as a Python package on Git Hub.1 1https://github.com/aidos-lab/mag_edge_pool available under a BSD 3-Clause License. ... We provide the code to reproduce the main experiments in the supplementary materials. |
| Open Datasets | Yes | All datasets are taken either from the TUDataset10 benchmark [45] or the Open Graph Benchmark11. ... 10https://chrsmrrs.github.io/datasets/ available under a CC-BY-4.0 license. 11https://ogb.stanford.edu/ available under an MIT licence. |
| Dataset Splits | Yes | We use 10-fold stratified cross-validation and further partition the training data into 90% training and 10% validation data while keeping the labels balanced between splits. ... we use three molecular datasets from the OGB benchmark with their predefined test, training, and validation splits [33]. |
| Hardware Specification | Yes | All experiments were conducted on a high-performance cluster with hardware specifications as detailed in Table S.1. In particular, all experiment were run requesting a single GPU with 32 GB video memory or less. ... Table S.1: Summary of the compute resources used for our experiments. Inventory Models Available CPUs Intel Xeon (Gold 6128, 6130, 6134, 6136, 6142, 6240, 6248R) Intel Xeon Platinum (8280L, 8480+, 8468, 8562Y+) Intel Xeon (E7-4850, E5620, 4114, 6126) AMD EPYC (7262, 7413, 7513, 7713, 7742) AMD Opteron (6128, 6164 HE, 6234, 6272, 6376 x2) Available GPUs NVIDIA Tesla (K80, P100, V100) NVIDIA A100 (20GB, 40GB, 80GB PCIe) NVIDIA H100 (80GB PCIe) NVIDIA Quadro RTX 8000 |
| Software Dependencies | Yes | The experiments reported in our study were implemented using spektral 1.3.1 [31]2, and tensorflow 2.16.2 [1]3. |
| Experiment Setup | Yes | The model includes pre-processing and post-processing MLPs with 2 layers, 256 hidden units, ReLU activation, and batch normalization. GNN(X, A) refers to a graph neural network layer, more specifically a general convolutional layer [57] with parameters chosen according to the best results achieved by You et al. [57]. ... Finally, we report the best test accuracy of each model trained using Adam with a cross-entropy loss (batch size 32, learning rate 0.0005, and early stopping based on the validation loss with a patience of 50 epochs). |