ML-GOOD: Towards Multi-Label Graph Out-Of-Distribution Detection

Authors: Tingyi Cai, Yunliang Jiang, Ming Li, Changqin Huang, Yi Wang, Qionghao Huang

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experimentation conducted on seven diverse sets of real-world multi-label graph datasets, encompassing crossdomain scenarios. The results show that the AUROC of MLGOOD is improved by 5.26% in intra-domain and 6.54% in cross-domain compared to the previous methods.
Researcher Affiliation Academia Tingyi Cai1,2, Yunliang Jiang2,1,3*, Ming Li4,2, Changqin Huang2, Yi Wang1,2, Qionghao Huang2 1School of Computer Science and Technology, Zhejiang Normal University, China 2Zhejiang Key Laboratory of Intelligent Education Technology and Application, Zhejiang Normal University, China 3School of Information Engineering, Huzhou University, China 4Zhejiang Institute of Optoelectronics, China EMAIL, EMAIL, EMAIL, EMAIL, EMAIL
Pseudocode Yes We summarize the complete algorithm for ML-GOOD in Appendix 1. 1The appendix is accessible via the provided Git Hub link.
Open Source Code Yes Code https://github.com/ca1man-2022/ML-GOOD
Open Datasets Yes In this experiment, we employ 6 real-world multilabel datasets including five molecular datasets, OGB-Proteins, (Hu et al. 2020), PPI, (Zitnik and Leskovec 2017), Hum Loc and Euk Loc (Zhao et al. 2023) and one citation network PCG.
Dataset Splits No To mitigate this, we conducted all experiments by randomly sampling 20% of this dataset, ensuring fairness and representativeness in comparisons. For datasets DBLP, PCG, Hum Loc, and Euk Loc, which are single-graph datasets lacking obvious domain information, we utilize feature interpolation, a method introduced by (Wu et al. 2023) for generating OOD data.
Hardware Specification Yes All experimental procedures are conducted on a NVIDIA RTX A6000 GPU device with 48 GB memory.
Software Dependencies No We uniformly use a 2-layer GCN (Kipf and Welling 2017) model as backbone encoder. We use the Adam optimizer (Kingma and Ba 2015) for optimization.
Experiment Setup Yes The weight decay is 0.01 and learning rate is 0.01. The results highlight the significant role margin hyperparameters mout and min in the effectiveness of ML-GOOD... impact of the mout { 60, 55, 45, 35, 25} and min { 85, 75, 65} margin hyperparameters on PPI; Right: impact of the weight hyperparameter λ on DBLP.