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 Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Fairness-aware Anomaly Detection via Fair Projection
Authors: Feng Xiao, Xiaoying Tang, Jicong Fan
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on real-world benchmarks demonstrate that our method achieves an improved trade-off between detection accuracy and fairness under both balanced and skewed data across different groups. Our experiments are conducted on six publicly available datasets from the literature on fairness in machine learning. |
| Researcher Affiliation | Academia | 1School of Science and Engineering, The Chinese University of Hong Kong (Shenzhen), Longgang, Shenzhen, Guangdong, 518172, P.R. China 2School of Data Science, The Chinese University of Hong Kong (Shenzhen), Longgang, Shenzhen, Guangdong, 518172, P.R. China |
| Pseudocode | No | The paper describes its methodology in textual form and through mathematical formulations. While it mentions reproducing code for a baseline using 'pseudo-code provided in the original paper [Zhang and Davidson, 2021]', it does not contain a dedicated pseudocode or algorithm block for its own proposed methods (Fair AD, Im-Fair AD, Ex-Fair AD). |
| Open Source Code | Yes | The source code is provided in supplementary materials. |
| Open Datasets | Yes | Our experiments are conducted on six publicly available datasets from the literature on fairness in machine learning [Zhang and Davidson, 2021, Chai and Wang, 2022, Han et al., 2023, Chen et al., 2024], where there are different kinds of sensitive information. The more detailed statistics of datasets and data splitting are provided in Appendix B.1. We both use standard (fairness-unaware) UAD methods, two-stage pipelines (FRL technique + UAD methods), and end-to-end fairness-aware UAD baselines. The specific download URL in Appendix B.1. |
| Dataset Splits | Yes | For fairness problems in unsupervised anomaly detection, the proportion of sample size across different demographic groups heavily influences the results [Meissen et al., 2023, Wu et al., 2024]. Therefore, in this work, we set balanced splitting and skewed splitting. The detailed splits for the training set and test set are provided in Tables 3 and 4. |
| Hardware Specification | Yes | ALL experiments were conducted on 20 Cores Intel(R) Xeon(R) Gold 6248 CPU with one NVIDIA Tesla V100 GPU, CUDA 12.0. |
| Software Dependencies | Yes | ALL experiments were conducted on 20 Cores Intel(R) Xeon(R) Gold 6248 CPU with one NVIDIA Tesla V100 GPU, CUDA 12.0. |
| Experiment Setup | Yes | Neural Network Architectures For the tabular datasets Adult, COMPAS, Credit, Titanic and SP, the neural networks for all methods are Multi-Layer Perceptrons (MLP). For the image dataset Celeb A, the neural networks used in all methods are Convolutional Neural Networks (CNN). We use Adam [Kingma and Ba, 2015] as the optimizer, and set coefficient α of entropy regularization term in the Sinkhorn distance to 0.1 in all experiments. We run each experiment five times and report the average results with standard variance. |