DiffGAD: A Diffusion-based Unsupervised Graph Anomaly Detector
Authors: Jinghan Li, Yuan Gao, Jinda Lu, Junfeng Fang, Congcong Wen, Hui Lin, Xiang Wang
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our comprehensive evaluation of Diff GAD, conducted on six real-world and large-scale datasets with various metrics, demonstrated its exceptional performance. Our code is available at: https://github.com/fortunato-all/Diff GAD. Extensive experiments over six real-world and large-scale datasets demonstrate our effectiveness, Theoretical and Empirical computational analysis illustrate our efficiency. |
| Researcher Affiliation | Academia | 1University of Science and Technology of China, 2New York University, 3China Academy of Electronics and Information Technology |
| Pseudocode | Yes | Algorithm 1 The training and inference procedure of Diff GAD. |
| Open Source Code | Yes | Our code is available at: https://github.com/fortunato-all/Diff GAD. |
| Open Datasets | Yes | Following the work in (Liu et al., 2022b) we employ 13 baselines as benchmarks on 6 real-world datasets (Weibo (Zhao et al., 2020), Reddit (Kumar et al., 2019; Wang et al., 2021), Disney (Sánchez et al., 2013), Books (Sánchez et al., 2013), Enron (Sánchez et al., 2013)), including a large-scale dataset Dgraph (Huang et al., 2022) for evaluation. |
| Dataset Splits | No | The paper references existing benchmark datasets and evaluation methodologies (Liu et al., 2022b) and mentions 'report the average performance with std results over 20 trials'. However, it does not explicitly provide the specific training, validation, and test split percentages or sample counts within its own text, relying on external references for dataset details. |
| Hardware Specification | Yes | The whole testing is conducted on a Linux server with a 2.90GHz Intel(R) Xeon(R) Platinum 8268 CPU, 1T RAM, and 1 Nvidia 2080 Ti GPU with 11GB memory. All the experiments were performed on a Linux server with a 3.00GHz Intel Xeon Gold 6248R CPU,1T RAM, and 1 NVIDIA A40 GPU with 45GB memory. |
| Software Dependencies | Yes | Environment. The key libraries and their versions used in the experiment are as follows: Python=3.11, CUDA_version=11.8, torch=2.0.1, pytorch_geometric=2.4.0, pygod=0.4.0, numpy=1.25.0 |
| Experiment Setup | Yes | Table 14: Hyper-parameter for different datasets. Hyper-paramter Type Weibo Reddit Disney Books Enron Dgraph Batch size AE FULL BATCH 8192 Epochs AE 300 Early stop AE No Dropout AE 0.3 0.3 0.3 0.1 0.1 0.3 Learning Rate AE 0.01 0.05 0.01 0.1 0.01 0.1 α AE 0.8 0.8 1.0 0.5 0.0 1.0 dimension AE 128 32 8 8 8 8 Batch size DM FULL BATCH 8192 Epochs DM 800 Early stop DM Yes Learning Rate DM 0.005 dimension DM 256 64 16 16 16 16 λ DM 1.0 0.8 2.0 2.0 2.0 1.0 |