Awesome

pyaad

Python version of Active Anomaly Discovery

Obsolete codebase -- no longer maintained

Please refer to (https://github.com/shubhomoydas/ad_examples) for current version.

Python libraries required:

numpy (1.13.3)
scipy (0.19.1)
scikit-learn (0.19.1)
cvxopt
pandas (0.21.0)
ranking
statsmodels
matplotlib (2.1.0)

Simplified Illustration of AAD

A much simplified version of AAD can be found in the project https://github.com/shubhomoydas/ad_examples. Please see the README.md for that project.

Running the code:

This codebase has four different algorithms:

• The LODA based AAD
• The Isolation Forest based AAD (does not support streaming incremental update)
• HS Trees based AAD (with streaming support)
• RS Forest based AAD (with streaming support)

To run the Isolation Forest / HS-Trees / RS-Forest based algorithms, the command has the following format:

bash ./tree_aad.sh <dataset> <budget> <reruns> <tau> <detector_type> <query_type> <query_confident[0|1]> <streaming[0|1]> <streaming_window>

for Isolation Forest, set <detector_type>=7; 
for HSTrees, set <detector_type>=11;
for RSForest, set <detector_type>=12;

example (with Isolation Forest, non-streaming):

bash ./tree_aad.sh toy2 35 1 0.03 7 1 0 0 512

Note: The above will generate 2D plots (tree partitions and score contours) under the 'temp' folder since <i>toy2</i> is a 2D dataset.

example (with HSTrees streaming):

bash ./tree_aad.sh toy2 35 1 0.03 11 1 0 1 256

Reference(s):

• Das, S., Wong, W-K., Dietterich, T., Fern, A. and Emmott, A. (2016). Incorporating Expert Feedback into Active Anomaly Discovery in the Proceedings of the IEEE International Conference on Data Mining. (http://web.engr.oregonstate.edu/~wongwe/papers/pdf/ICDM2016.AAD.pdf) (https://github.com/shubhomoydas/aad/blob/master/overview/ICDM2016-AAD.pptx)

• Das, S., Wong, W-K., Fern, A., Dietterich, T. and Siddiqui, A. (2017). Incorporating Feedback into Tree-based Anomaly Detection, KDD Interactive Data Exploration and Analytics (IDEA) Workshop. (http://poloclub.gatech.edu/idea2017/papers/p25-das.pdf) (https://github.com/shubhomoydas/pyaad/blob/master/presentations/IDEA17_slides.pptx)

Running the original (non-tree/LODA based) AAD Code:

For the most straightforward execution of the code, assume that we have the original datafile and another file that has anomaly scores from an ensemble of detectors. One example of these files (and their formats) can be found under the folder 'sampledata'.

The output will have two files: file '-baseline.csv' shows the number of true anomalies detected with each iteration if we do not incorporate feedback; and the file '-num_seen.csv' shows the number of true anomalies detected when we incorporate feedback.

Sample run command (AAD original (non-tree/LODA based)):

python ./pyalad/alad.py --startcol=2 --labelindex=1 --header --randseed=42 --dataset=toy --datafile=./sampledata/toy.csv --scoresfile=./sampledata/toy_scores.csv --querytype=1 --detector_type=3 --constrainttype=1 --sigma2=0.5 --reps=1 --reruns=1 --budget=10 --tau=0.03 --Ca=100 --Cn=1 --Cx=1000 --withprior --unifprior --runtype=simple --log_file=./temp/pyaad.log --resultsdir=./temp --ensembletype=regular --debug

R Version

An older implementation in R is available at: https://github.com/shubhomoydas/aad