/ADBench

Official Implement of "ADBench: Anomaly Detection Benchmark".

Primary LanguagePythonBSD 2-Clause "Simplified" LicenseBSD-2-Clause

Official code and data repository of ADBench: Anomaly Detection Benchmark (NeurIPS 2022). Please star, watch, and fork ADBench for the active updates!

Citing ADBench:

Our ADBench benchmark paper is now available on arxiv and OpenReview. If you find this work useful or use some our released datasets, we would appreciate citations to the following paper:

@inproceedings{han2022adbench,  
      title={ADBench: Anomaly Detection Benchmark},   
      author={Songqiao Han and Xiyang Hu and Hailiang Huang and Mingqi Jiang and Yue Zhao},  
      booktitle={Neural Information Processing Systems (NeurIPS)}
      year={2022},  
}

Who Are We? ✨

ADBench is a collaborative effort between researchers at Shanghai University of Finance and Economics (SUFE) and Carnegie Mellon University (CMU). The project is designed and conducted by Minqi Jiang (SUFE) and Yue Zhao (CMU), and Xiyang Hu (CMU) --the author(s) of important anomaly detection libraries, including anomaly detection for tabular (PyOD), time-series (TODS), and graph data (PyGOD).

Why Do You Need ADBench?

ADBench is (to our best knowledge) the most comprehensive tabular anomaly detection benchmark, where we analyze the performance of 30 anomaly detection algorithms on 57 datasets (where we introduced 10 new datasets). By analyzing both research needs and deployment requirements in industry, ADBench conducts 98,436 experiments with three major angles:

  1. the effect of supervision (e.g., ground truth labels) by including 14 unsupervised, 7 semi-supervised, and 9 supervised methods;
  2. algorithm performance under different types of anomalies by simulating the environments with 4 types of anomalies; and
  3. algorithm robustness and stability under 3 settings of data corruptions.

Key Takeaways in 1 Minute:

  1. ‼️ surprisingly none of the benchmarked unsupervised algorithms is statistically better than others, emphasizing the importance of algorithm selection;
  2. ‼️ with merely 1% labeled anomalies, most semi-supervised methods can outperform the best unsupervised method, justifying the importance of supervision;
  3. in controlled environments, we observe that best unsupervised methods for specific types of anomalies are even better than semi- and fully-supervised methods, revealing the necessity of understanding data characteristics;
  4. semi-supervised methods show potential in achieving robustness in noisy and corrupted data, possibly due to their efficiency in using labels and feature selection;
  5. ⁉️ and many more can be found in our papers (Section 4)

The Figure below provides an overview of our proposed ADBench (see our paper for details).

ADBench


How to use ADBench?

We envision three primary usages of ADBench:

  • Have better understanding of anomaly detection algorithms: please read our paper for details.
  • Conduct future research on anomaly detection: we list 4 important future research questions in the paper--see Section 4 to see some thoughts!
  • Access rich algorithm implementation and datasets: see details below for how to use them
  • Benchmark your anomaly detection algorithms: see notebook for instruction.

Dependency

The experiment code is written in Python 3 and built on a number of Python packages:

  • scikit-learn==0.20.3
  • pyod==1.0.0
  • Keras==2.8.0
  • tensorflow==2.8.0
  • torch==1.9.0
  • rtdl==0.0.13
  • delu
  • lightgbm
  • xgboost
  • catboost
  • copulas

Quickly implement ADBench for benchmarking AD algorithms.

We present the following example for quickly implementing ADBench in three different Angles illustrated in the paper. Currently, 57 datasets can be used for evaluating 30 algorithms in ADBench, and we encourage to test your customized datasets/algorithms in our ADBench testbed.

Angle I: Availability of Ground Truth Labels (Supervision)

from data_generator import DataGenerator
from myutils import Utils

# one can use our already included datasets
data_generator = DataGenerator(dataset='6_cardio.npz')
# specify the ratio of labeled anomalies to generate X and y
# la could be any float number in [0.0, 1.0]
data = data_generator.generator(la=0.1) 

# or specify the X and y of your customized data
# data_generator = DataGenerator(dataset=None)
# data = data_generator.generator(X=X, y=y, la=0.1)

# import AD algorithms (e.g., DevNet) and initialization
from baseline.DevNet.run import DevNet
model = DevNet(seed=42)

# fitting
model.fit(X_train=data['X_train'], y_train=data['y_train'])

# prediction
score = model.predict_score(data['X_test'])

# evaluation
utils = Utils()
result = utils.metric(y_true=data['y_test'], y_score=score)

Angle II: Types of Anomalies

# For Angle II, different types of anomalies are generated as the following
data_generator = DataGenerator(dataset='6_cardio.npz')
# the type of anomalies could be 'local', 'global', 'dependency' or 'cluster'.
data = data_generator.generator(realistic_synthetic_mode='local')

Angle III: Model Robustness with Noisy and Corrupted Data

# For Angle III, different data noises and corruptions are added as the following
data_generator = DataGenerator(dataset='6_cardio.npz')
# the type of anomalies could be 'duplicated_anomalies', 'irrelevant_features' or 'label_contamination'.
data = data_generator.generator(noise_type='duplicated_anomalies')
  • We also provide an example for quickly implementing ADBench, as shown in notebook.
  • For complete results of ADBench, please refer to our paper.
  • For reproduce experiment results of ADBench, please run the code.

Datasets

ADBench includes 57 datasets, as shown in the following Table.

  • Among them, 47 widely-used real-world datasets are gathered for model evaluation, which cover many application domains, including healthcare (e.g., disease diagnosis), audio and language processing (e.g., speech recognition), image processing (e.g., object identification), finance (e.g., financial fraud detection), etc.

  • we introduce 10 more complex datasets from CV and NLP domains with more samples and richer features in ADBench. Pretrained models are applied to extract data embedding from NLP and CV datasets to access more complex representation. Please see the datasets folder and our paper for detailed information.

  • We organize the above 57 datasets into user-friendly format. All the datasets are named as "number_data.npz" in the datasets folder. For example, one can evaluate AD algorithms on the cardio dataset by the following codes. For multi-class dataset like CIFAR10, additional class numbers should be specified as "number_data_class.npz". Please see the folder for more details.

  • We provide the data processing code for NLP datasets and for CV datasets in Google Colab, where one can quickly reproduce our procedures via the free GPUs. We hope this could be helpful for the AD community.

We have unified all the datasets in .npz format, and you can read a dataset by the following script

import numpy as np
data = np.load('6_cardio.npz', allow_pickle=True)
X, y = data['X'], data['y']
Number Data # Samples # Features # Anomaly % Anomaly Category
1 ALOI 49534 27 1508 3.04 Image
2 annthyroid 7200 6 534 7.42 Healthcare
3 backdoor 95329 196 2329 2.44 Network
4 breastw 683 9 239 34.99 Healthcare
5 campaign 41188 62 4640 11.27 Finance
6 cardio 1831 21 176 9.61 Healthcare
7 Cardiotocography 2114 21 466 22.04 Healthcare
8 celeba 202599 39 4547 2.24 Image
9 census 299285 500 18568 6.20 Sociology
10 cover 286048 10 2747 0.96 Botany
11 donors 619326 10 36710 5.93 Sociology
12 fault 1941 27 673 34.67 Physical
13 fraud 284807 29 492 0.17 Finance
14 glass 214 7 9 4.21 Forensic
15 Hepatitis 80 19 13 16.25 Healthcare
16 http 567498 3 2211 0.39 Web
17 InternetAds 1966 1555 368 18.72 Image
18 Ionosphere 351 32 126 35.90 Oryctognosy
19 landsat 6435 36 1333 20.71 Astronautics
20 letter 1600 32 100 6.25 Image
21 Lymphography 148 18 6 4.05 Healthcare
22 magic.gamma 19020 10 6688 35.16 Physical
23 mammography 11183 6 260 2.32 Healthcare
24 mnist 7603 100 700 9.21 Image
25 musk 3062 166 97 3.17 Chemistry
26 optdigits 5216 64 150 2.88 Image
27 PageBlocks 5393 10 510 9.46 Document
28 pendigits 6870 16 156 2.27 Image
29 Pima 768 8 268 34.90 Healthcare
30 satellite 6435 36 2036 31.64 Astronautics
31 satimage-2 5803 36 71 1.22 Astronautics
32 shuttle 49097 9 3511 7.15 Astronautics
33 skin 245057 3 50859 20.75 Image
34 smtp 95156 3 30 0.03 Web
35 SpamBase 4207 57 1679 39.91 Document
36 speech 3686 400 61 1.65 Linguistics
37 Stamps 340 9 31 9.12 Document
38 thyroid 3772 6 93 2.47 Healthcare
39 vertebral 240 6 30 12.50 Biology
40 vowels 1456 12 50 3.43 Linguistics
41 Waveform 3443 21 100 2.90 Physics
42 WBC 223 9 10 4.48 Healthcare
43 WDBC 367 30 10 2.72 Healthcare
44 Wilt 4819 5 257 5.33 Botany
45 wine 129 13 10 7.75 Chemistry
46 WPBC 198 33 47 23.74 Healthcare
47 yeast 1484 8 507 34.16 Biology
48 CIFAR10 5263 512 263 5.00 Image
49 FashionMNIST 6315 512 315 5.00 Image
50 MNIST-C 10000 512 500 5.00 Image
51 MVTec-AD See Table B2. Image
52 SVHN 5208 512 260 5.00 Image
53 Agnews 10000 768 500 5.00 NLP
54 Amazon 10000 768 500 5.00 NLP
55 Imdb 10000 768 500 5.00 NLP
56 Yelp 10000 768 500 5.00 NLP
57 20newsgroups See Table B3. NLP

Algorithms

Compared to the previous benchmark studies, we have a larger algorithm collection with

  1. latest unsupervised AD algorithms like DeepSVDD and ECOD;
  2. SOTA semi-supervised algorithms, including DeepSAD and DevNet;
  3. latest network architectures like ResNet in computer vision (CV) and Transformer in natural language processing (NLP) domain ---we adapt ResNet and FTTransformer models for tabular AD in the proposed ADBench; and
  4. ensemble learning methods like LightGBM, XGBoost, and CatBoost. The Figure below shows the algorithms (14 unsupervised, 7 semi-supervised, and 9 supervised algorithms) in ADBench. Algorithms

For each algorithm, we also introduce its specific implementation in the following Table. The only thing worth noting is that model name should be specified (especially for those models deployed by their corresponding package, e.g., PyOD). The following codes show the example to import AD models. Please see the Table for complete AD models included in ADBench and their import methods.

from baseline.PyOD import PYOD
model = PYOD(model_name='XGBOD') # initialization
model.fit(X_train, y_train) # fit
score = model.predict_score(X_test) # predict
Model Year Type DL Import Source
PCA Before 2017 Unsup from baseline.PyOD import PYOD Link
OCSVM Before 2017 Unsup from baseline.PyOD import PYOD Link
LOF Before 2017 Unsup from baseline.PyOD import PYOD Link
CBLOF Before 2017 Unsup from baseline.PyOD import PYOD Link
COF Before 2017 Unsup from baseline.PyOD import PYOD Link
HBOS Before 2017 Unsup from baseline.PyOD import PYOD Link
KNN Before 2017 Unsup from baseline.PyOD import PYOD Link
SOD Before 2017 Unsup from baseline.PyOD import PYOD Link
COPOD 2020 Unsup from baseline.PyOD import PYOD Link
ECOD 2022 Unsup from baseline.PyOD import PYOD Link
IForest† Before 2017 Unsup from baseline.PyOD import PYOD Link
LODA† Before 2017 Unsup from baseline.PyOD import PYOD Link
DeepSVDD 2018 Unsup from baseline.PyOD import PYOD Link
DAGMM 2018 Unsup from baseline.DAGMM.run import DAGMM Link
GANomaly 2018 Semi from baseline.GANomaly.run import GANomaly Link
XGBOD† 2018 Semi from baseline.PyOD import PYOD Link
DeepSAD 2019 Semi from baseline.DeepSAD.src.run import DeepSAD Link
REPEN 2018 Semi from baseline.REPEN.run import REPEN Link
DevNet 2019 Semi from baseline.DevNet.run import DevNet Link
PReNet 2020 Semi from baseline.PReNet.run import PReNet /
FEAWAD 2021 Semi from baseline.FEAWAD.run import FEAWAD Link
NB Before 2017 Sup from baseline.Supervised import supervised Link
SVM Before 2017 Sup from baseline.Supervised import supervised Link
MLP Before 2017 Sup from baseline.Supervised import supervised Link
RF† Before 2017 Sup from baseline.Supervised import supervised Link
LGB† 2017 Supervised from baseline.Supervised import supervised Link
XGB† Before 2017 Sup from baseline.Supervised import supervised Link
CatB† 2019 Sup from baseline.Supervised import supervised Link
ResNet 2019 Sup from baseline.FTTransformer.run import FTTransformer Link
FTTransformer 2019 Sup from baseline.FTTransformer.run import FTTransformer Link
  • '†' marks ensembling. This symbol is not included in the model name.
  • Un-, semi-, and fully-supervised methods are denoted as unsup, semi and sup, respectively.