Below is one of the projects I did in my RA at Uchicago computer science department that is open to public. Many others entail private data and private algirithms that are build for paper publishing. They are not made public until the paper is reviewed and published in the journal.
Python TimeSeries Anomaly Detection (PTSA)
PTSA(in deveopment) is a comprehensive and scalable Python toolkit for detecting anomalies in timeseries data. I already included 20 algorithms in the package, and more will come as I keep developing the package. There is a python notebook demo about how to use the package.
API Demo:
# train the Polynomial detector
from ptsa.models.poly import Poly
model = Poly(window = 200, power = 0)
model.fit(data)
# get outlier scores
from ptsa.models.distance import Fourier
measure = Fourier()
measure.detector = model
measure.set_param()
model.decision_function(measure=measure)It is recommended to use pip for installation. The required package is in requirement.txt.
git clone https://github.com/alacrity2001/AnomalyDetection
cd code/ptsa
pip install .Required Dependencies:
- Python 3.5, 3.6, or 3.7
- combo>=0.0.8
- joblib
- numpy>=1.13
- numba>=0.35
- pandas>=0.25
- scipy>=0.19.1
- scikit_learn>=0.19.1
- six
- statsmodels
- suod
Optional Dependencies (see details below):
- keras (optional, required for AutoEncoder, ANN, LSTM)
- matplotlib (optional, required for running examples)
- pandas (optional, required for reading demo)
- tensorflow (optional, required for AutoEncoder, other backend works)
- pmdarima (optional, required for ARIMA)
- arch (optional, required for ARCH measure)
- tsfresh (optional, required for tf feature extraction)
- hurst (optional, tf feature extraction)
(Note) Individual Detection Algorithms :
The ptsa/models has all the scripts.
The machine learning models for iforest, cblof, cof, knn, loci, lof, mcd, pca, ocsvm are which I modifed from PYOD libaray. The RPIF.py is hybrid isolation forest which is in https://github.com/pfmarteau/HIF
I coded all the others includind the feature transformation, dissimilarity measure, etc. All the dissimilarity measure is quite standard, except SSA, which I referenced from the paper https://bourbon.usc.edu/iml/projects/nsf-dc-0917340/anomaly.pdf
utils.utility are some functions I gathered from sklearn library and other authors of other alogirthms.
ptsa.others.normats.norma is the norma algorithm whose author didn't pusblish his paper yet. The code is directly copied from their folder with some debuggings for some errors encountered when applying our data.
ptsa.utils.metrics has the class metricor I coded as a convenient tool to evaluate algorithms. The script I coded to compute ranged recall/precision is referenced from the paper https://arxiv.org/pdf/1803.03639.pdf