GDeR: Safeguarding Efficiency, Balancing, and Robustness via Prototypical Graph Pruning

Authors: Guibin Zhang, Haonan Dong, yuchen zhang, Zhixun Li, Dingshuo Chen, Kai Wang, Tianlong Chen, Yuxuan Liang, Dawei Cheng, Kun Wang

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on five datasets across three GNN backbones, demonstrate that GDe R (I) achieves or surpasses the performance of the full dataset with 30% 50% fewer training samples, (II) attains up to a 2.81 lossless training speedup, and (III) outperforms state-of-the-art pruning methods in imbalanced training and noisy training scenarios by 0.3% 4.3% and 3.6% 7.8%, respectively.
Researcher Affiliation Collaboration Guibin Zhang 1,2, Haonan Dong 1, Yuchen Zhang2, Zhixun Li3, Dingshuo Chen4, Kai Wang5, Tianlong Chen6, Yuxuan Liang7, Dawei Cheng 1,2, Kun Wang 8 1Tongji Univerity, 2Shanghai AI Laboratory, 3CUHK, 4UCAS, 5NUS, 6UNC-Chapel Hill, 7HKUST (Guangzhou) 8NTU
Pseudocode Yes Algorithm 1: Algorithm workflow of GDe R
Open Source Code Yes The source code is available at https://github.com/ins1stenc3/GDe R.
Open Datasets Yes We test GDe R on two widely-used datasets, MUTAG [38] and DHFR [87]; two OGB large-scale datasets, OGBG-MOLHIV and OGBG-MOLPBCA [88]; one large-scale chemical compound dataset ZINC [89].
Dataset Splits Yes Following [40], we adopt a 25%/25%/50% train/validation/test random split for the MUTAG and DHFR under imbalanced scenarios and 80%/10%/10% under normal and biased scenarios, both reporting results across 20 data splits.
Hardware Specification Yes All the experiments are conducted on NVIDIA Tesla V100 (32GB GPU), using Py Torch and Py Torch Geometric framework.
Software Dependencies No The paper mentions 'Py Torch and Py Torch Geometric framework' but does not specify their version numbers.
Experiment Setup Yes The hyperparameters in GDe R include the temperature coefficient τ, prototype count K, loss-specific coefficient λ1 and λ2. Practically, we uniformly set K = 2, and tune the other three by grid searching: τ {1e 3, 1e 4, 1e 5}, λ {1e 1, 5e 1},λ {1e 1, 1e 5}.