Towards Self-Interpretable Graph-Level Anomaly Detection

Authors: Yixin Liu, Kaize Ding, Qinghua Lu, Fuyi Li, Leo Yu Zhang, Shirui Pan

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on 16 datasets demonstrate the anomaly detection capability and self-interpretability of SIGNET.
Researcher Affiliation Academia 1Monash University, 2Northwestern University, 3Data61, CSIRO, 4Northwest A&F University, 5The University of Adelaide, 6Griffith University
Pseudocode Yes More discussion about methodology, including the pseudo-code algorithm of SIGNET, the comparison between SIGNET and existing method, and the complexity analysis of SIGNET, is illustrated in Appendix E.
Open Source Code Yes Our code is available at https://github.com/yixinliu233/SIGNET.
Open Datasets Yes We also verify the anomaly detection performance of SIGNET on 10 TU datasets [58], following the setting in [4]. ... MNIST-0 and MNIST-1 are two GLAD datasets derived from MNIST-75sp superpixel dataset [59]. ... MUTAG is a molecular property prediction dataset [61].
Dataset Splits No Given the training set Gtr that contains a number of normal graphs, we aim at learning an explainable GLAD model f : G (R, G) that is able to predict the abnormality of a graph and provide corresponding explanations. In specific, given a graph Gi from the test set Gte with normal and abnormal graphs... It mentions "training set" and "test set" but does not explicitly provide validation split percentages or counts.
Hardware Specification No The paper does not provide specific hardware details such as GPU/CPU models or detailed computer specifications used for running experiments.
Software Dependencies No In SIGNET, we use GIN [2] and Hyper-Conv [30] as the GNN and HGNN encoders. The paper names software components but does not provide specific version numbers for any software dependencies.
Experiment Setup Yes For all methods, we perform 5 random runs and report the average performance. We use Adam optimizer with learning rate 0.001 and weight decay 0.0001. We train 100 epochs with early stopping (patience 20). We use a batch size of 128. For SIGNET, the GNN encoder (GIN) and HGNN encoder (Hyper-Conv) are 2-layer, and the hidden dimension is 64. The bottleneck subgraph extractor (MLP) has 2 layers. We set the temperature parameter τ = 0.5. The number of negative samples is 128 for Info-NCE. The trade-off parameter β = 0.001.