This repository contains the supplementary material as well as the implementation of our paper Unsupervised Surrogate Anomaly Detection.
In supplementary.pdf we provide the equivalents of the AUC-ROC performance plots from our experimental evaluation for the AUC-PR metric, as well as the detailed results for the specific datasets both for AUC-ROC and AUC-PR.
The supplementary also covers additional analysis results for DEAN with regard to the effect from incorporation of a learnable shift (bias term) in the network architecture. We further demonstrate the adaptability of DEAN based on the inclusion of fairness criteria into the predictions.
dean.py implements the DEAN method.
main.py allows to directly conduct the training and evaluation of a DEAN ensemble as specified in the configuration file config.yaml for a given dataset as provided in the data folder.
Alternatively, the configuration parameters may also be overwritten using command line arguments, e.g.:
python main.py --dataset data/Cardio.npz --model_count 5
A suitable conda environment based on the requirements specified in env.yaml may be created via:
conda env create -f env.yaml
The folder competitors contains information regarding the implementation and parametrization of the competitor algorithms used during the experimental evaluation.
The folder fairness contains modifications of DEAN to for a proof-of-concept to include fairness criteria in the predictions, as referred to in the supplementary.