[ICDM 2020] Meta-AAD: Active Anomaly Detection with Deep Reinforcement Learning

This is the implementation of ICDM 2020 paper Meta-AAD: Active Anomaly Detection with Deep Reinforcement Learning. We propose to learn a meta-policy with deep reinforcement learning to optimize the performance of active anomaly detection. Please refer the paper for more deteails.

📢 Do you want to learn more about data labeling? Please check out our data-centric AI survey and data-centric AI resources!

Cite this Work

If you find this project helpful, please cite

@inproceedings{DBLP:conf/icdm/ZhaLWH20,
  author    = {Daochen Zha and
               Kwei{-}Herng Lai and
               Mingyang Wan and
               Xia Hu},
  editor    = {Claudia Plant and
               Haixun Wang and
               Alfredo Cuzzocrea and
               Carlo Zaniolo and
               Xindong Wu},
  title     = {Meta-AAD: Active Anomaly Detection with Deep Reinforcement Learning},
  booktitle = {20th {IEEE} International Conference on Data Mining, {ICDM} 2020,
               Sorrento, Italy, November 17-20, 2020},
  pages     = {771--780},
  publisher = {{IEEE}},
  year      = {2020},
  url       = {https://doi.org/10.1109/ICDM50108.2020.00086},
  doi       = {10.1109/ICDM50108.2020.00086},
  timestamp = {Wed, 17 Feb 2021 11:24:58 +0100},
  biburl    = {https://dblp.org/rec/conf/icdm/ZhaLWH20.bib},
  bibsource = {dblp computer science bibliography, https://dblp.org}
}

Installation

Make sure you have python 3.5+ installed.

git clone https://github.com/daochenzha/Meta-AAD.git
cd Meta-AAD
pip install -r requirments.txt
pip install -e .

Training a Meta-Policy

Train a meta-policy with train.py. The important arguments are as follows.

• --train: the datasets used for training, seperated by commas.
• --test: the datasets used for testing, seperated by commas.
• --num_timesteps: the number of training steps of reinforcement learning agents
• --log: where the log and models will be outputted

By default, the reinforcement learning training log will be saved in log/, the anomaly discovery curves will be saved in log/anomaly_curves, and the trained model will be saved in log/.

Evaluating a Trained Model

You may evaluate a trained model with evaluate.py. The important arguments are as follows.

• --load: the path to model.zip file.
• --test: the datasets used for testing, seperated by commas.

Baselines

We provide two baselines in this repo for comparison: a random query strategy and IForest query strategy. They are available in evaluate_baselines.py. For other baselines, please refer to the following repos.

• AAD: https://github.com/shubhomoydas/ad_examples
• FIF: https://github.com/siddiqmd/FeedbackIsolationForest
• SSDO: https://github.com/Vincent-Vercruyssen/anomatools