Bundle Neural Network for message diffusion on graphs

Authors: Jacob Bamberger, Federico Barbero, Xiaowen Dong, Michael Bronstein

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

Reproducibility Variable Result LLM Response
Research Type Experimental We support our claims with formal analysis and synthetic experiments. Empirically, Bu NNs perform strongly on heterophilic and long-range tasks, which demonstrates their robustness on a diverse range of challenging real-world tasks. In summary, our contributions are the following: [...] We show that Bu NNs perform well on heterophilic and long-range tasks, for instance, achieving a new state-of-the-art result on the Peptides-func dataset (Section 6.2).
Researcher Affiliation Collaboration Jacob Bamberger1 Federico Barbero1 Xiaowen Dong1 Michael Bronstein1,2 1University of Oxford 2AITHYRA EMAIL
Pseudocode Yes Algorithm 1 Taylor expansion implementation of a Bu NN layer [...] Algorithm 2 Spectral implementation of a Bu NN layer
Open Source Code Yes All code can be found at https://github.com/jacobbamberger/Bu NN
Open Datasets Yes We evaluate Bu NNs on the Long Range Graph Benchmark (Dwivedi et al., 2022) and the heterophilic tasks from Platonov et al. (2023).
Dataset Splits Yes We use 100 samples for both training and testing. [...] For peptides-func and peptides-struct we use a fixed parameter budget of 500k as in Dwivedi et al. (2022). We fix hyper-parameters to be the best GCN hyper-parameters from T onshoff et al. (2023) [...] For the heterophilic graphs we use the source code from Platonov et al. (2023) in which we add our layer definition.
Hardware Specification Yes All experiments were ran on a cluster using NVIDIA A10 (24 GB) GPUs, each experiment using at most 1 GPU. Each machine in the cluster has 64 cores of Intel(R) Xeon(R) Gold 6326 CPU at 2.90GHz, and 500GB of RAM available.
Software Dependencies No We use the Pytorch package from (?). [...] We use Adam optimizer with 10 3 learning rate.
Experiment Setup Yes Dropout 0.1 [...] Optimizer: Adam W with a cosine annealing learning rate schedule and linear warmup. [...] Batch size 200 [...] For the heterophilic graphs we use [...] GELU activation functions, the Adam optimizer with learning rate 3 10 5, and train all models for 2000 epochs and select the best epoch based on the validation set. Tables 5, 6, and 7 provide detailed hyperparameter grids and best values.