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].
Enhancing Fairness in Unsupervised Graph Anomaly Detection through Disentanglement
Authors: Wenjing Chang, Kay Liu, Philip S. Yu, Jianjun Yu
TMLR 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirical evaluations on real-world datasets reveal that DEFEND performs effectively in GAD and significantly enhances fairness compared to state-of-the-art baselines. |
| Researcher Affiliation | Academia | Wenjing Chang EMAIL Computer Network Information Center, Chinese Academy of Sciences University of Chinese Academy of Sciences Kay Liu EMAIL University of Illinois Chicago Philip S. Yu EMAIL University of Illinois Chicago Jianjun Yu EMAIL Computer Network Information Center, Chinese Academy of Sciences |
| Pseudocode | Yes | A Pseudo Code The overall training process of DEFEND is presented in Algorithm 1. Firstly, given an attributed graph G, we aim to train a variational graph autoencoder, in which the disentangled encoder fe can separate sensitiveirrelevant representations while maintaining information for reconstructing G by the decoder fa and fx (Line 4-6). We encourage the disentanglement with a learnable adversary gω optimized by Ladv (Line 8-9). The optimization of fe, fa and fx and that of gω are conducted adversarially. Next, the frozen fe captures sensitive-irrelevant node representations (Line 14), and the decoder fϕ solely reconstructs attributes (Line 15). We utilize reconstruction error as the anomaly score and constrain the correlation between reconstruction error and predicted sensitive attributes (Lines 16-19). Finally, only fϕ will be updated by Lad in the anomaly detection phase. |
| Open Source Code | Yes | Our code is available at https://github.com/AhaChang/DEFEND. |
| Open Datasets | Yes | Datasets. We employ three real-world datasets for fair GAD, which provide both real sensitive attributes and ground-truth labels for GAD. In Reddit and Twitter (Neo et al., 2024) datasets, the sensitive attribute is the political leaning of users, while the anomaly label is assigned to misinformation spreaders. The Credit (Agarwal et al., 2021) dataset focuses on payment default detection, with age as the sensitive attribute. Details of these datasets are summarized in Table 1. |
| Dataset Splits | No | The paper lists dataset statistics in Table 1 (Nodes, Edges, Attributes, ratios of groups/anomalies) and mentions using real-world datasets, but does not provide explicit training, validation, or test splits (e.g., percentages, sample counts, or references to predefined splits). |
| Hardware Specification | Yes | We conduct all experiments on one Linux server with an NVIDIA TESLA A800 GPU (80 GB RAM). |
| Software Dependencies | No | For DOMINANT, Co LA, CONAD and VGOD, we use the code and default hyper-parameters provided by Py GOD1 (Liu et al., 2022c). For Fair Walk2 and EDITS3, we implement them using the code published by their authors. For Fair OD, HIN, and Correlation, we implement the code provided by Fair GAD4. The paper mentions PyGOD and other frameworks but does not provide specific version numbers for these software dependencies or programming languages/libraries like Python or PyTorch. |
| Experiment Setup | Yes | We employ the Adam optimizer with a learning rate set to 0.001 for Reddit and Credit datasets and 0.005 for Twitter. We set the maximum training epoch in disentangled representation learning as 100, and adopt an early stopping strategy when the loss does not decrease for 20 epochs. In the anomaly detection phase, we train the decoder fϕ for 100 epochs. We tune α and γ in Equation 15 from {0.1, 0.5, 1.0, 1.5, 2.0, 2.5}, and the weight of correlation constrains β in Equation 19 from {1e 15, 5e 15, 1e 10, 5e 10, 1e 9}, respectively. |