Nonparametric Teaching for Graph Property Learners
Authors: Chen Zhang, Weixin Bu, Zeyi Ren, Zhengwu Liu, Yik Chung Wu, Ngai Wong
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Lastly, we conduct extensive experiments to validate the effectiveness of Gra NT in a range of scenarios, covering both graph-level and node-level tasks. Our key contributions are listed as follows: We demonstrate the effectiveness of Gra NT through extensive experiments in graph property learning, covering regression and classification at both graph and node levels. Specifically, Gra NT saves training time for graph-level regression (-36.62%), graph-level classification (-38.19%), node-level regression (-30.97%) and node-level classification (-47.30%), while upkeeping its generalization performance. The overall results on the test set are shown in Table 1, which clearly highlights the effectiveness of Gra NT in graph property learning |
| Researcher Affiliation | Collaboration | 1Department of Electrical and Electronic Engineering, The University of Hong Kong, HKSAR, China 2Reversible Inc. Correspondence to: Chen Zhang <EMAIL>, Ngai Wong <EMAIL>. |
| Pseudocode | Yes | Algorithm 1 Gra NT Algorithm Input: Target mapping f realized by a dense set of graphproperty pairs, initial GCN fθ0, the size of selected training set m N, small constant ϵ > 0 and maximal iteration number T. |
| Open Source Code | No | Our project page is available at https://chen2hang.github.io/_publications/nonparametric_teaching_for_graph_proerty_learners/grant.html. This is a project page and does not explicitly state that the source code for the methodology is released, nor does it provide a direct link to a code repository. |
| Open Datasets | Yes | We evaluate Gra NT using several widely recognized benchmark datasets as follows: QM9 (Wu et al., 2018): 130k organic molecules graphs with quantum chemical properties (regression task); ZINC (Gómez-Bombarelli et al., 2018): 250k molecular graphs with bioactivity and solubility chemical properties (regression task); ogbg-molhiv (Hu et al., 2020): 41k molecular graphs with HIV inhibitory activity properties (binary classification task); ogbg-molpcba (Hu et al., 2020): 438k molecular graphs with bioactivity properties (multi-task binary classification task). |
| Dataset Splits | Yes | Dataset Splitting. The train / val / test split configurations for the benchmark datasets are provided in Table 5. Table 5: Dataset splitting for the benchmark datasets. QM9 110000 10000 10831 ZINC 220011 24445 5000 ogbg-molhiv 32901 4113 4113 ogbg-molpcba 350343 43793 43793 gen-reg 30000 10000 10000 gen-cls 30000 10000 10000 |
| Hardware Specification | Yes | Device Setup. We mainly conduct experiments using NVIDIA Geforce RTX 3090 (24G). Furthermore, on the AMD Instinct MI210 (64GB) device, we also validate the effectiveness of Gra NT for graph-level tasks on the QM9 dataset, highlighting its cross-device effectiveness. |
| Software Dependencies | No | The paper mentions the use of GCNs and references techniques like 'Reduce LROn Plateau', but it does not specify concrete version numbers for any software libraries, programming languages, or other ancillary software components used in the experiments. |
| Experiment Setup | Yes | Hyperparameter Settings. The key hyperparameter settings for all benchmark datasets are listed in Table 6. Table 6: Key hyperparameter settings for the benchmark datasets, with the start-ratio specified for Gra NT. Dataset lr κ-list batch-size start-ratio (Gra NT) epochs QM9 0.00005 [3, 2] 256 0.05 750 ZINC 0.0004 [5, 4, 2, 2] 256 0.05 1000 ogbg-molhiv 0.01 [4, 3, 2, 2] 500 0.1 600 ogbg-molpcba 0.015 [5, 4, 3, 2, 2] 128 0.1 800 gen-reg 0.0002 [3, 2] 100 0.05 250 gen-cls 0.0002 [4, 3] 200 0.05 500 |