Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].

CurvGAD: Leveraging Curvature for Enhanced Graph Anomaly Detection

Authors: Karish Grover, Geoffrey J. Gordon, Christos Faloutsos

ICML 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experimentation over 10 real-world datasets (both homophilic and heterophilic) demonstrates an improvement of up to 6.5% over state-of-the-art GAD methods. The code is available at: https://github.com/ karish-grover/curvgad.
Researcher Affiliation Academia Karish Grover 1 Geoffrey J. Gordon 1 Christos Faloutsos 1 1Carnegie Mellon University. Correspondence to: Karish Grover <EMAIL>.
Pseudocode Yes Algorithm 1 Discrete Ollivier-Ricci Flow Algorithm 2 Product manifold signature estimation
Open Source Code Yes The code is available at: https://github.com/ karish-grover/curvgad.
Open Datasets Yes We evaluate Curv GAD on 10 datasets, each containing organic node-level anomalies, to assess its effectiveness across homophilic and heterophilic settings. These datasets span multiple domains, including social media, e-commerce, and financial networks. Specifically, (1) Weibo (Zhao et al., 2020; Liu et al., 2022), (2) Reddit (Kumar et al., 2019; Liu et al., 2022), (3) Questions (Platonov et al., 2023), and (4) T-Social (Tang et al., 2022)...
Dataset Splits Yes If splits are not provided, we adopt the strategy from (Tang et al., 2022), partitioning nodes into 40%/20%/40% for training, validation, and testing, as detailed in Table 6. To ensure robustness, we perform ten random splits per dataset and report the average performance.
Hardware Specification Yes All experiments are conducted on A6000 GPUs (48GB), using a total manifold dimension of d P = 48, a learning rate of 0.01, and a filterbank comprising F = 8 filters.
Software Dependencies No The paper mentions: "we leverage the Îș-stereographic product manifold implementation from Geoopt1 and use Riemannian Adam for gradient-based learning across product manifolds." and footnote 1: "https://github.com/geoopt/geoopt". However, specific version numbers for Geoopt or other software dependencies are not provided.
Experiment Setup Yes All experiments are conducted on A6000 GPUs (48GB), using a total manifold dimension of d P = 48, a learning rate of 0.01, and a filterbank comprising F = 8 filters. Appendix E.2 enlists the hyperparameter configurations tried and we analyse the time complexity of Curv GAD in Appendix C. Table 7: Hyperparameter Tuning Configurations Description F {3, 5, 8, 10, 20, 25} Total number of graph filters. ... lr {1e 4, 2e 3, 0.001, 0.01} Learning rate weight decay {0, 1e 4, 5e 4} Weight decay