Revisiting Score Propagation in Graph Out-of-Distribution Detection

Authors: Longfei Ma, Yiyou Sun, Kaize Ding, Zemin Liu, Fei Wu

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

Reproducibility Variable Result LLM Response
Research Type Experimental Empirical evaluations affirm the superiority of our proposed method, outperforming strong OOD detection baselines in various scenarios and settings.
Researcher Affiliation Academia Longfei Ma1 , Yiyou Sun2 , Kaize Ding3, Zemin Liu1 , Fei Wu1 1Zhejiang University, 2University of Wisconsin-Madison, 3Northwestern University
Pseudocode No The paper describes algorithmic steps and procedures in text but does not include formal pseudocode blocks or figures explicitly labeled 'Algorithm' or 'Pseudocode'.
Open Source Code Yes To ensure reproducibility, we have made our code and relevant data publicly available at https://github.com/longfei-ma/GRASP.
Open Datasets Yes Datasets. We conduct extensive experiments using 10 real-world datasets that span diverse domains, scales, and structures (homophily or heterophily). A high-level summary of the dataset statistics is provided in Table 1, with a detailed information of the datasets and the comprehensive description of ID/OOD split in Appendix C. Specifically, Cora [61] serves as a widely recognized citation network.
Dataset Splits No The paper defines labeled (Vl) and unlabeled (Vu) node sets, treating unlabeled nodes as test ID (Vuid) or test OOD (Vuood). While hyperparameters are tuned via sensitivity analysis (Appendix D.4), the paper does not explicitly describe a separate validation dataset split with specific percentages or counts.
Hardware Specification No The paper provides computational complexity analysis, including time and memory costs, for the algorithms and baselines in Appendix D.6 (Table 16, Table 17). However, it does not specify the exact hardware (e.g., CPU/GPU models, memory configurations) on which these experiments were conducted.
Software Dependencies No The paper mentions using specific GNN architectures like 'Graph Convolutional Network (GCN)' and references 'Pytorch Geometric package', but it does not provide specific version numbers for these or other key software dependencies required for replication.
Experiment Setup Yes All pre-trained models possess a layer depth of 2. ... We set the propagation number k as 8, with percentile values α = 5 and β = 50. The sensitivity analysis of the hyperparameters is included in Appendix D.4.