kaist-dmlab / ARCUS

Adaptive Model Pooling for Online Deep Anomaly Detection from a Complex Evolving Data Stream

This is the implementation of ARCUS published in KDD 2022 [paper]

Required packages

• Tensorflow 2.2.0
• Python 3.8.3
• Scikit-learn 0.23.1
• Numpy 1.18.5
• Pandas 1.0.5

Data sets description and link

• The last column in each data set file refers to the anomaly label (1: anomaly, 0:normal)
  
• Data sets link
  
• The link includes the small data sets (also included in the repository) and large data sets exceeding 100MB
  

How to run ARCUS

Parameters

• model_type: type of model, one of ["RAPP", "RSRAE", "DAGMM"]
• inf_type: type of inference, one of ["INC", "ADP"] where "INC" for incremental and "ADP" for adaptive (proposed)
• batch_size: batch size (default: 512)
• min_batch_size: min batch size (default: 32)
• init_epoch: initial number of epochs for creating models (default: 5)
• intm_epoch: interim number of epochs for training models after initialization (default: 1)
• hidden_dim: latent dimensionality of AE (default: the number of pricipal component explaining at least 70% of variance)
• layer_num: the number of layers in AE

Training script

$ python main.py --model_type RAPP --dataset_name MNIST_AbrRec --inf_type ADP --batch_size 512 --min_batch_size 32 --init_epoch 5 --intm_epoch 1 --hidden_dim 24 --layer_num 3 --learning_rate 1e-4 --reliability_thred 0.95 --similarity_thred 0.80 --seed 42 --gpu '0' 
----------------------------
Data set: MNIST_AbrRec
Model type:  RAPP
AUC: 0.909

Example concept drift adaptation of ARCUS in INSECTS data sets

Default model layer size (learning rate) used for ARCUS

5. Citation

@inproceedings{yoon2022arcus,
  title={Adaptive Model Pooling for Online Deep Anomaly Detection from a Complex Evolving Data Stream},
  author={Yoon, Susik, and Lee, Youngjun, and Lee, Jae-Gil and Lee, Byung Suk},
  booktitle={Proceedings of the 28th ACM SIGKDD International Conference on Knowledge Discovery \& Data Mining},
  pages={--},
  year={2022}
}