A minimal and pure python fork of @nicolasmiller's library pyculiarity. That is a Python port of Twitter's AnomalyDetection R Package. The original source and examples are available here.
The wdm00006fork is focused on python3 compatibility and no dependency on R. This is done using statsmodel's young tsa.seasonal_decompose, which differs in output from the Loess STL implementation used by the original pyculiarity library. The results are not identical as a result, but are pretty close.
wdm00006 also stripped out some unused/unimplimented code to try to make this a little more readable/understandable. Another important distinction of the wdm00006 fork is that the timestamp must be in unix timestamp format.
The original library is on pypi as pyculiarity. Since I (hokiegeek2) am interesting in testing both the original and this version of pyculiarity, I changed the package name to pyculiar. To install this version of pyculiarity (again, named pyculiar), run the following command from the root project directory:
python2/3 setup.py install
As in Twitter's package, there are two top level functions, one for time-series data and one for simple vector processing, detect_ts and detect_vec respectively. The first one expects a two-column Pandas DataFrame consisting of timestamps and values. The second expects either a single-column DataFrame or a Series.
Here's an example of loading Twitter's example data (included in the tests directory) with Pandas and passing it to Pyculiarity for processing.
hokiegeek2 NOTE: The only_last parameter is not supported in this version of pycularity. In addition, the raw_data.csv file bundled with this project cannot be processed with this version of pyculiarity since the timestamp must be in unix timestamp format.
from pyculiarity import detect_ts
import matplotlib.pyplot as plt
import pandas as pd
import matplotlib
matplotlib.style.use('ggplot')
__author__ = 'willmcginnis'
if __name__ == '__main__':
# first run the models
twitter_example_data = pd.read_csv('../tests/raw_data.csv', usecols=['timestamp', 'count'])
results = detect_ts(twitter_example_data, max_anoms=0.05, alpha=0.001, direction='both')
# format the twitter data nicely
twitter_example_data['timestamp'] = pd.to_datetime(twitter_example_data['timestamp'])
twitter_example_data.set_index('timestamp', drop=True)
# make a nice plot
f, ax = plt.subplots(2, 1, sharex=True)
ax[0].plot(twitter_example_data['timestamp'], twitter_example_data['count'], 'b')
ax[0].plot(results['anoms'].index, results['anoms']['anoms'], 'ro')
ax[0].set_title('Detected Anomalies')
ax[1].set_xlabel('Time Stamp')
ax[0].set_ylabel('Count')
ax[1].plot(results['anoms'].index, results['anoms']['anoms'], 'b')
ax[1].set_ylabel('Anomaly Magnitude')
plt.show()
Which will give the plot:
The tests are run with nose as follows:
nosetests .
Changes Copyright 2016 Will McGinnis Python port Copyright 2015 Nicolas Steven Miller Original R source Copyright 2015 Twitter, Inc and other contributors
Licensed under the GPLv3